9 Agent Orchestration

Introduction

Agentic AI systems coordinate multiple specialized agents to solve complex tasks beyond single-model capabilities. As AI applications evolve from simple chatbots to autonomous systems capable of executing multi-step workflows, understanding agent orchestration becomes essential for building production-grade AI solutions.

信息

The Agentic Shift: 2024-2025 marked the transition from "prompt engineering" to "agent engineering" - designing systems where LLMs act as reasoning engines orchestrating tools, memory, and other agents.

Why Multi-Agent Systems?

Single-agent limitations:

• Context window constraints: Complex tasks exceed token limits
• Cognitive overload: One agent handling diverse roles reduces quality
• Specialization trade-offs: General capability vs. domain expertise
• Error propagation: Single point of failure affects entire workflow

Multi-agent advantages:

• Divide and conquer: Decompose complex problems into manageable subtasks
• Specialization: Each agent optimized for specific role
• Parallel execution: Faster processing through concurrent operations
• Resilience: Failures isolated to individual agents
• Scalability: Add agents without redesigning system

Agent Architecture Fundamentals

┌─────────────────────────────────────────────────────────────────────┐
│                        Agent Architecture                           │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌─────────────┐    ┌─────────────┐    ┌─────────────────────────┐ │
│  │   Persona   │    │    Goal     │    │       Backstory         │ │
│  │  (Role)     │    │  (Objective)│    │  (Context/Expertise)    │ │
│  └─────────────┘    └─────────────┘    └─────────────────────────┘ │
│         │                 │                       │                 │
│         └─────────────────┼───────────────────────┘                 │
│                           ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐  │
│  │                    Reasoning Engine (LLM)                     │  │
│  │  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────────┐  │  │
│  │  │ Planning │  │ Reasoning│  │ Decision │  │ Self-Critique│  │  │
│  │  └──────────┘  └──────────┘  └──────────┘  └──────────────┘  │  │
│  └──────────────────────────────────────────────────────────────┘  │
│                           │                                         │
│         ┌─────────────────┼─────────────────┐                      │
│         ▼                 ▼                 ▼                       │
│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐              │
│  │    Tools    │   │   Memory    │   │Communication│              │
│  │  (Actions)  │   │  (State)    │   │  (Messages) │              │
│  └─────────────┘   └─────────────┘   └─────────────┘              │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

---

Core Concepts

Agent Definition Framework

Component	Description	Example
Agent	Autonomous unit with role, goal, and backstory	"Senior Security Analyst with 10 years experience"
Tool	External capability (API, database, function)	Web search, code execution, database query
Memory	State persistence across interactions	Conversation history, learned facts, task progress
Crew/Team	Collection of agents assembled for tasks	Research team, code review committee
Process	Orchestration pattern (sequential, hierarchical, parallel)	Assembly line, management hierarchy

Spring AI Agent Configuration

@Configuration
public class AgentConfiguration {

    @Bean
    public AgentDefinition researcherAgent(ChatClient.Builder builder) {
        return AgentDefinition.builder()
            .name("researcher")
            .role("Senior Research Analyst")
            .goal("Gather comprehensive, accurate information from multiple sources")
            .backstory("""
                You are a meticulous researcher with expertise in academic literature,
                market analysis, and technical documentation. You verify facts from
                multiple sources and clearly distinguish between established facts
                and speculation. You have a PhD in Information Science and 15 years
                of experience in competitive intelligence.
                """)
            .chatClient(builder
                .defaultSystem(generateSystemPrompt())
                .build())
            .tools(List.of(
                webSearchTool(),
                documentRetrieverTool(),
                citationManagerTool()
            ))
            .build();
    }

    @Bean
    public AgentDefinition writerAgent(ChatClient.Builder builder) {
        return AgentDefinition.builder()
            .name("writer")
            .role("Technical Content Writer")
            .goal("Transform research into clear, engaging, actionable content")
            .backstory("""
                You are an award-winning technical writer who excels at making
                complex topics accessible. You have written for major tech
                publications and authored several O'Reilly books. You balance
                technical accuracy with readability.
                """)
            .chatClient(builder.build())
            .tools(List.of(
                grammarCheckerTool(),
                readabilityAnalyzerTool()
            ))
            .build();
    }
}

@Data
@Builder
public class AgentDefinition {
    private String name;
    private String role;
    private String goal;
    private String backstory;
    private ChatClient chatClient;
    private List<FunctionCallback> tools;
    private MemoryStore memory;
}

---

1. Sequential Orchestration

The Assembly Line: Agents execute in a linear chain where each agent's output becomes the next agent's input.

Architecture

┌──────────────────────────────────────────────────────────────────────┐
│                    Sequential Pipeline                               │
├──────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  Input ──▶ [Parser] ──▶ [Enricher] ──▶ [Analyzer] ──▶ [Formatter]   │
│              │            │              │              │            │
│              ▼            ▼              ▼              ▼            │
│           Tokens       Metadata       Insights       Output          │
│                                                                      │
│  Flow: Strictly linear, output(N) = input(N+1)                      │
│  Error: Pipeline halts on failure                                    │
│  State: Passed forward through context                               │
│                                                                      │
└──────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class SequentialOrchestrator {

    private final List<AgentDefinition> pipeline;
    private final MeterRegistry registry;

    public SequentialOrchestrator(
            @Qualifier("parserAgent") AgentDefinition parser,
            @Qualifier("enricherAgent") AgentDefinition enricher,
            @Qualifier("analyzerAgent") AgentDefinition analyzer,
            @Qualifier("formatterAgent") AgentDefinition formatter,
            MeterRegistry registry) {
        this.pipeline = List.of(parser, enricher, analyzer, formatter);
        this.registry = registry;
    }

    public PipelineResult process(String input) {
        return process(input, new PipelineContext());
    }

    public PipelineResult process(String input, PipelineContext context) {
        String currentOutput = input;
        List<StageResult> stageResults = new ArrayList<>();

        Timer.Sample pipelineTimer = Timer.start(registry);

        for (int i = 0; i < pipeline.size(); i++) {
            AgentDefinition agent = pipeline.get(i);
            String stageName = agent.getName();

            log.info("Stage {}/{}: {} processing", i + 1, pipeline.size(), stageName);

            Timer.Sample stageTimer = Timer.start(registry);

            try {
                StageResult result = executeStage(agent, currentOutput, context);

                stageTimer.stop(registry.timer("pipeline.stage",
                    "stage", stageName, "status", "success"));

                stageResults.add(result);
                currentOutput = result.output();

                // Update context with stage metadata
                context.addStageMetadata(stageName, result.metadata());

                log.info("Stage {} completed: {} tokens in, {} tokens out",
                    stageName, result.inputTokens(), result.outputTokens());

            } catch (Exception e) {
                stageTimer.stop(registry.timer("pipeline.stage",
                    "stage", stageName, "status", "error"));

                log.error("Stage {} failed: {}", stageName, e.getMessage());

                return PipelineResult.builder()
                    .success(false)
                    .failedStage(stageName)
                    .stageResults(stageResults)
                    .error(e.getMessage())
                    .build();
            }
        }

        pipelineTimer.stop(registry.timer("pipeline.total", "status", "success"));

        return PipelineResult.builder()
            .success(true)
            .output(currentOutput)
            .stageResults(stageResults)
            .context(context)
            .build();
    }

    private StageResult executeStage(AgentDefinition agent, String input,
                                     PipelineContext context) {
        String systemPrompt = buildSystemPrompt(agent, context);

        ChatResponse response = agent.getChatClient().prompt()
            .system(systemPrompt)
            .user(u -> u.text("""
                Previous context: {context}

                Current input to process:
                {input}

                Execute your role and provide output for the next stage.
                """)
                .param("context", context.getSummary())
                .param("input", input))
            .call()
            .chatResponse();

        return StageResult.builder()
            .stageName(agent.getName())
            .output(response.getResult().getOutput().getContent())
            .inputTokens(response.getMetadata().getUsage().getPromptTokens())
            .outputTokens(response.getMetadata().getUsage().getCompletionTokens())
            .metadata(extractMetadata(response))
            .build();
    }

    private String buildSystemPrompt(AgentDefinition agent, PipelineContext context) {
        return String.format("""
            You are a %s.

            Role: %s
            Goal: %s

            Background:
            %s

            Pipeline Position: You are part of a sequential processing pipeline.
            Your output will be passed to the next stage for further processing.

            Output Requirements:
            - Provide clear, structured output
            - Include any metadata needed by subsequent stages
            - Flag any issues or uncertainties
            """,
            agent.getName(),
            agent.getRole(),
            agent.getGoal(),
            agent.getBackstory()
        );
    }
}

@Data
@Builder
public class PipelineResult {
    private boolean success;
    private String output;
    private String failedStage;
    private String error;
    private List<StageResult> stageResults;
    private PipelineContext context;
}

@Data
@Builder
public class StageResult {
    private String stageName;
    private String output;
    private int inputTokens;
    private int outputTokens;
    private Map<String, Object> metadata;
}

Advanced: Resumable Pipeline with Checkpoints

@Service
public class ResumablePipeline {

    private final SequentialOrchestrator orchestrator;
    private final CheckpointStore checkpointStore;

    public PipelineResult processWithCheckpoints(String jobId, String input) {
        // Check for existing checkpoint
        Optional<Checkpoint> checkpoint = checkpointStore.getCheckpoint(jobId);

        if (checkpoint.isPresent()) {
            log.info("Resuming from checkpoint: stage {} for job {}",
                checkpoint.get().getLastCompletedStage(), jobId);
            return resumeFromCheckpoint(jobId, checkpoint.get());
        }

        // Fresh start with checkpoint saving
        return processWithCheckpointSaving(jobId, input);
    }

    private PipelineResult processWithCheckpointSaving(String jobId, String input) {
        PipelineContext context = new PipelineContext();
        context.setJobId(jobId);

        // Register checkpoint callback
        context.setOnStageComplete((stageName, output) -> {
            checkpointStore.saveCheckpoint(Checkpoint.builder()
                .jobId(jobId)
                .lastCompletedStage(stageName)
                .stageOutput(output)
                .context(context)
                .timestamp(Instant.now())
                .build());
        });

        return orchestrator.process(input, context);
    }

    private PipelineResult resumeFromCheckpoint(String jobId, Checkpoint checkpoint) {
        PipelineContext context = checkpoint.getContext();
        return orchestrator.processFromStage(
            checkpoint.getLastCompletedStage(),
            checkpoint.getStageOutput(),
            context
        );
    }
}

Best For

Use Case	Example
Data Processing Pipelines	ETL workflows, log processing
Document Workflows	Intake → Classification → Extraction → Validation
Content Transformation	Raw data → Analysis → Report → Summary
Compliance Processing	Check A → Check B → Check C → Approval

---

2. Hierarchical Orchestration

The Russian Doll: A supervisor agent decomposes tasks and delegates to specialist sub-agents, collecting and synthesizing results.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    Hierarchical Architecture                        │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│                    ┌─────────────────────┐                         │
│                    │   Supervisor Agent  │                         │
│                    │  (Task Decomposer)  │                         │
│                    └──────────┬──────────┘                         │
│                               │                                     │
│               ┌───────────────┼───────────────┐                    │
│               │               │               │                     │
│               ▼               ▼               ▼                     │
│    ┌──────────────┐  ┌──────────────┐  ┌──────────────┐           │
│    │  Researcher  │  │   Analyst    │  │    Writer    │           │
│    │   Agent      │  │    Agent     │  │    Agent     │           │
│    └──────────────┘  └──────────────┘  └──────────────┘           │
│           │                 │                 │                     │
│           │    ┌────────────┘                 │                     │
│           │    │                              │                     │
│           ▼    ▼                              ▼                     │
│    ┌─────────────────────────────────────────────────┐             │
│    │              Results Aggregation                 │             │
│    └─────────────────────────────────────────────────┘             │
│                               │                                     │
│                               ▼                                     │
│                    ┌─────────────────────┐                         │
│                    │   Supervisor Agent  │                         │
│                    │   (Synthesizer)     │                         │
│                    └─────────────────────┘                         │
│                               │                                     │
│                               ▼                                     │
│                         Final Output                                │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class HierarchicalOrchestrator {

    private final AgentDefinition supervisor;
    private final Map<String, AgentDefinition> specialists;
    private final ExecutorService executorService;
    private final MeterRegistry registry;

    public HierarchicalOrchestrator(
            @Qualifier("supervisorAgent") AgentDefinition supervisor,
            Map<String, AgentDefinition> specialists,
            MeterRegistry registry) {
        this.supervisor = supervisor;
        this.specialists = specialists;
        this.executorService = Executors.newFixedThreadPool(
            Math.min(specialists.size(), 10)
        );
        this.registry = registry;
    }

    public OrchestratedResult execute(String task) {
        log.info("Starting hierarchical execution for task: {}",
            task.substring(0, Math.min(100, task.length())));

        // Phase 1: Supervisor decomposes task
        TaskDecomposition decomposition = decomposeTask(task);
        log.info("Task decomposed into {} subtasks", decomposition.getSubtasks().size());

        // Phase 2: Delegate to specialists (parallel or sequential based on dependencies)
        Map<String, SubtaskResult> results = executeSubtasks(decomposition);

        // Phase 3: Supervisor synthesizes results
        String synthesis = synthesizeResults(task, decomposition, results);

        return OrchestratedResult.builder()
            .task(task)
            .decomposition(decomposition)
            .subtaskResults(results)
            .synthesis(synthesis)
            .build();
    }

    private TaskDecomposition decomposeTask(String task) {
        String availableSpecialists = specialists.keySet().stream()
            .map(name -> "- " + name + ": " + specialists.get(name).getGoal())
            .collect(Collectors.joining("\n"));

        return supervisor.getChatClient().prompt()
            .system(s -> s.text("""
                You are a project manager who decomposes complex tasks into subtasks.

                Available specialists:
                {specialists}

                For the given task:
                1. Break it into specific, actionable subtasks
                2. Assign each subtask to the most appropriate specialist
                3. Identify dependencies between subtasks
                4. Specify the execution order (parallel where possible)

                Output as JSON matching the TaskDecomposition schema.
                """)
                .param("specialists", availableSpecialists))
            .user(task)
            .call()
            .entity(TaskDecomposition.class);
    }

    private Map<String, SubtaskResult> executeSubtasks(TaskDecomposition decomposition) {
        Map<String, SubtaskResult> results = new ConcurrentHashMap<>();

        // Group subtasks by execution phase (respecting dependencies)
        List<List<Subtask>> executionPhases = groupByDependencies(decomposition);

        for (int phase = 0; phase < executionPhases.size(); phase++) {
            List<Subtask> parallelTasks = executionPhases.get(phase);
            log.info("Executing phase {}/{} with {} parallel tasks",
                phase + 1, executionPhases.size(), parallelTasks.size());

            // Execute all tasks in this phase in parallel
            List<CompletableFuture<SubtaskResult>> futures = parallelTasks.stream()
                .map(subtask -> CompletableFuture.supplyAsync(
                    () -> executeSubtask(subtask, results),
                    executorService
                ))
                .collect(Collectors.toList());

            // Wait for all tasks in phase to complete
            CompletableFuture.allOf(futures.toArray(new CompletableFuture[0])).join();

            // Collect results
            for (int i = 0; i < parallelTasks.size(); i++) {
                SubtaskResult result = futures.get(i).join();
                results.put(parallelTasks.get(i).getId(), result);
            }
        }

        return results;
    }

    private SubtaskResult executeSubtask(Subtask subtask,
                                         Map<String, SubtaskResult> previousResults) {
        AgentDefinition specialist = specialists.get(subtask.getAssignedAgent());

        if (specialist == null) {
            log.warn("No specialist found for: {}, using supervisor",
                subtask.getAssignedAgent());
            specialist = supervisor;
        }

        // Gather dependency outputs
        String dependencyContext = subtask.getDependencies().stream()
            .filter(previousResults::containsKey)
            .map(dep -> String.format("From %s:\n%s", dep,
                previousResults.get(dep).getOutput()))
            .collect(Collectors.joining("\n\n"));

        Timer.Sample timer = Timer.start(registry);

        try {
            String output = specialist.getChatClient().prompt()
                .system(s -> s.text("""
                    Role: {role}
                    Goal: {goal}

                    Background:
                    {backstory}

                    Execute the assigned subtask thoroughly and provide a complete response.
                    """)
                    .param("role", specialist.getRole())
                    .param("goal", specialist.getGoal())
                    .param("backstory", specialist.getBackstory()))
                .user(u -> u.text("""
                    Subtask: {subtask}

                    Context from previous subtasks:
                    {context}

                    Execute this subtask and provide detailed output.
                    """)
                    .param("subtask", subtask.getDescription())
                    .param("context", dependencyContext.isEmpty() ? "None" : dependencyContext))
                .call()
                .content();

            timer.stop(registry.timer("hierarchical.subtask",
                "agent", subtask.getAssignedAgent(), "status", "success"));

            return SubtaskResult.builder()
                .subtaskId(subtask.getId())
                .agent(subtask.getAssignedAgent())
                .output(output)
                .success(true)
                .build();

        } catch (Exception e) {
            timer.stop(registry.timer("hierarchical.subtask",
                "agent", subtask.getAssignedAgent(), "status", "error"));

            return SubtaskResult.builder()
                .subtaskId(subtask.getId())
                .agent(subtask.getAssignedAgent())
                .success(false)
                .error(e.getMessage())
                .build();
        }
    }

    private String synthesizeResults(String originalTask,
                                     TaskDecomposition decomposition,
                                     Map<String, SubtaskResult> results) {
        String resultsFormatted = decomposition.getSubtasks().stream()
            .map(subtask -> {
                SubtaskResult result = results.get(subtask.getId());
                return String.format("""
                    ## Subtask: %s
                    Assigned to: %s
                    Status: %s

                    Output:
                    %s
                    """,
                    subtask.getDescription(),
                    subtask.getAssignedAgent(),
                    result.isSuccess() ? "Completed" : "Failed: " + result.getError(),
                    result.isSuccess() ? result.getOutput() : "N/A"
                );
            })
            .collect(Collectors.joining("\n---\n"));

        return supervisor.getChatClient().prompt()
            .system("""
                You are synthesizing results from multiple specialist agents.

                Your role:
                1. Integrate insights from all subtask outputs
                2. Resolve any conflicts or inconsistencies
                3. Fill gaps where subtasks may have missed important aspects
                4. Create a coherent, comprehensive final response
                5. Highlight key findings and recommendations
                """)
            .user(u -> u.text("""
                Original Task:
                {task}

                Subtask Results:
                {results}

                Synthesize these into a comprehensive final response that fully
                addresses the original task.
                """)
                .param("task", originalTask)
                .param("results", resultsFormatted))
            .call()
            .content();
    }

    private List<List<Subtask>> groupByDependencies(TaskDecomposition decomposition) {
        // Topological sort to group tasks by execution phase
        Map<String, Set<String>> dependencyGraph = new HashMap<>();
        Map<String, Subtask> subtaskMap = new HashMap<>();

        for (Subtask subtask : decomposition.getSubtasks()) {
            subtaskMap.put(subtask.getId(), subtask);
            dependencyGraph.put(subtask.getId(), new HashSet<>(subtask.getDependencies()));
        }

        List<List<Subtask>> phases = new ArrayList<>();
        Set<String> completed = new HashSet<>();

        while (completed.size() < subtaskMap.size()) {
            List<Subtask> currentPhase = new ArrayList<>();

            for (Map.Entry<String, Set<String>> entry : dependencyGraph.entrySet()) {
                String taskId = entry.getKey();
                Set<String> deps = entry.getValue();

                if (!completed.contains(taskId) && completed.containsAll(deps)) {
                    currentPhase.add(subtaskMap.get(taskId));
                }
            }

            if (currentPhase.isEmpty() && completed.size() < subtaskMap.size()) {
                // Circular dependency detected - break by taking any remaining task
                String remaining = subtaskMap.keySet().stream()
                    .filter(id -> !completed.contains(id))
                    .findFirst()
                    .orElseThrow();
                currentPhase.add(subtaskMap.get(remaining));
            }

            phases.add(currentPhase);
            currentPhase.forEach(t -> completed.add(t.getId()));
        }

        return phases;
    }
}

@Data
@Builder
public class TaskDecomposition {
    private List<Subtask> subtasks;
    private String executionStrategy;
    private Map<String, Object> metadata;
}

@Data
@Builder
public class Subtask {
    private String id;
    private String description;
    private String assignedAgent;
    private List<String> dependencies;
    private int priority;
}

Best For

Use Case	Example
Complex Project Management	Product launches, research projects
Research and Analysis	Market research, competitive analysis
Multi-stage Workflows	Document processing with multiple specialists
Enterprise Applications	Customer onboarding, loan processing

---

3. Parallel Orchestration

The Octopus: Multiple agents execute simultaneously on the same input, with results aggregated.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    Parallel Architecture                            │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│                           Input                                     │
│                             │                                       │
│           ┌─────────────────┼─────────────────┐                    │
│           │                 │                 │                     │
│           ▼                 ▼                 ▼                     │
│    ┌──────────────┐  ┌──────────────┐  ┌──────────────┐           │
│    │   Security   │  │    Style     │  │  Performance │           │
│    │   Analyst    │  │   Reviewer   │  │   Engineer   │           │
│    └──────────────┘  └──────────────┘  └──────────────┘           │
│           │                 │                 │                     │
│           │   ┌─────────────┼─────────────┐   │                    │
│           │   │             │             │   │                     │
│           ▼   ▼             ▼             ▼   ▼                     │
│    ┌──────────────────────────────────────────────────┐            │
│    │              Result Aggregator                    │            │
│    │    (Merge, Deduplicate, Prioritize Findings)     │            │
│    └──────────────────────────────────────────────────┘            │
│                             │                                       │
│                             ▼                                       │
│                    Synthesized Report                               │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class ParallelOrchestrator {

    private final Map<String, AgentDefinition> agents;
    private final AgentDefinition synthesizer;
    private final ExecutorService executorService;
    private final MeterRegistry registry;

    public ParallelOrchestrator(
            @Qualifier("parallelAgents") Map<String, AgentDefinition> agents,
            @Qualifier("synthesizerAgent") AgentDefinition synthesizer,
            MeterRegistry registry) {
        this.agents = agents;
        this.synthesizer = synthesizer;
        this.executorService = Executors.newCachedThreadPool();
        this.registry = registry;
    }

    public ParallelResult execute(String input) {
        return execute(input, agents.keySet());
    }

    public ParallelResult execute(String input, Set<String> selectedAgents) {
        log.info("Starting parallel execution with {} agents", selectedAgents.size());

        Timer.Sample totalTimer = Timer.start(registry);

        // Launch all agents in parallel
        Map<String, CompletableFuture<AgentResult>> futures = selectedAgents.stream()
            .filter(agents::containsKey)
            .collect(Collectors.toMap(
                name -> name,
                name -> CompletableFuture.supplyAsync(
                    () -> executeAgent(name, input),
                    executorService
                )
            ));

        // Wait for all with timeout
        CompletableFuture<Void> allFutures = CompletableFuture.allOf(
            futures.values().toArray(new CompletableFuture[0])
        );

        try {
            allFutures.get(5, TimeUnit.MINUTES);
        } catch (TimeoutException e) {
            log.warn("Some agents timed out after 5 minutes");
        } catch (Exception e) {
            log.error("Error waiting for agents", e);
        }

        // Collect results (completed or failed)
        Map<String, AgentResult> results = new HashMap<>();
        for (Map.Entry<String, CompletableFuture<AgentResult>> entry : futures.entrySet()) {
            try {
                results.put(entry.getKey(), entry.getValue().getNow(
                    AgentResult.builder()
                        .agentName(entry.getKey())
                        .success(false)
                        .error("Timed out")
                        .build()
                ));
            } catch (Exception e) {
                results.put(entry.getKey(), AgentResult.builder()
                    .agentName(entry.getKey())
                    .success(false)
                    .error(e.getMessage())
                    .build());
            }
        }

        // Synthesize results
        String synthesis = synthesizeResults(input, results);

        totalTimer.stop(registry.timer("parallel.total"));

        return ParallelResult.builder()
            .input(input)
            .agentResults(results)
            .synthesis(synthesis)
            .completedCount((int) results.values().stream().filter(AgentResult::isSuccess).count())
            .totalCount(results.size())
            .build();
    }

    private AgentResult executeAgent(String agentName, String input) {
        AgentDefinition agent = agents.get(agentName);
        Timer.Sample timer = Timer.start(registry);

        log.info("Agent {} starting analysis", agentName);

        try {
            ChatResponse response = agent.getChatClient().prompt()
                .system(s -> s.text("""
                    Role: {role}
                    Goal: {goal}

                    Background:
                    {backstory}

                    You are part of a parallel review team. Focus exclusively on your
                    area of expertise. Other specialists are handling other aspects.

                    Provide findings in this structure:
                    1. Executive Summary (2-3 sentences)
                    2. Detailed Findings (prioritized list)
                    3. Recommendations (actionable items)
                    4. Risk Assessment (if applicable)
                    """)
                    .param("role", agent.getRole())
                    .param("goal", agent.getGoal())
                    .param("backstory", agent.getBackstory()))
                .user(input)
                .call()
                .chatResponse();

            timer.stop(registry.timer("parallel.agent",
                "agent", agentName, "status", "success"));

            return AgentResult.builder()
                .agentName(agentName)
                .output(response.getResult().getOutput().getContent())
                .success(true)
                .tokensUsed(response.getMetadata().getUsage().getTotalTokens())
                .build();

        } catch (Exception e) {
            timer.stop(registry.timer("parallel.agent",
                "agent", agentName, "status", "error"));

            log.error("Agent {} failed: {}", agentName, e.getMessage());

            return AgentResult.builder()
                .agentName(agentName)
                .success(false)
                .error(e.getMessage())
                .build();
        }
    }

    private String synthesizeResults(String input, Map<String, AgentResult> results) {
        String allResults = results.entrySet().stream()
            .map(e -> String.format("""
                ## %s Review
                Status: %s

                %s
                """,
                e.getKey(),
                e.getValue().isSuccess() ? "Completed" : "Failed: " + e.getValue().getError(),
                e.getValue().isSuccess() ? e.getValue().getOutput() : "N/A"
            ))
            .collect(Collectors.joining("\n---\n"));

        return synthesizer.getChatClient().prompt()
            .system("""
                You are synthesizing reviews from multiple specialist agents.

                Your responsibilities:
                1. Merge overlapping findings
                2. Resolve conflicting assessments (explain the conflict)
                3. Prioritize combined recommendations
                4. Create executive summary of all perspectives
                5. Highlight consensus vs. divergent opinions
                """)
            .user(u -> u.text("""
                Original Input:
                {input}

                Agent Reviews:
                {reviews}

                Create a unified report that synthesizes all perspectives.
                """)
                .param("input", input.substring(0, Math.min(1000, input.length())))
                .param("reviews", allResults))
            .call()
            .content();
    }
}

// Specialized Code Review Implementation
@Service
public class ParallelCodeReviewer extends ParallelOrchestrator {

    @Bean
    public Map<String, AgentDefinition> parallelAgents(ChatClient.Builder builder) {
        return Map.of(
            "security", createSecurityAgent(builder),
            "style", createStyleAgent(builder),
            "performance", createPerformanceAgent(builder),
            "architecture", createArchitectureAgent(builder),
            "testing", createTestingAgent(builder)
        );
    }

    private AgentDefinition createSecurityAgent(ChatClient.Builder builder) {
        return AgentDefinition.builder()
            .name("security")
            .role("Security Analyst")
            .goal("Identify security vulnerabilities, injection risks, and unsafe patterns")
            .backstory("""
                You are a senior security engineer with OSCP, CISSP certifications.
                You specialize in application security and have found vulnerabilities
                in major open source projects. You focus on OWASP Top 10, CWE,
                and secure coding practices.
                """)
            .chatClient(builder.build())
            .build();
    }

    private AgentDefinition createPerformanceAgent(ChatClient.Builder builder) {
        return AgentDefinition.builder()
            .name("performance")
            .role("Performance Engineer")
            .goal("Identify performance bottlenecks, memory issues, and optimization opportunities")
            .backstory("""
                You are a performance optimization expert with experience scaling
                systems to millions of users. You focus on time complexity,
                memory allocation, caching strategies, and database query optimization.
                """)
            .chatClient(builder.build())
            .build();
    }

    // ... other agent definitions
}

Best For

Use Case	Example
Code Review	Security + Style + Performance in parallel
Research	Multiple sources investigated simultaneously
Content Generation	Multiple creative variations
Risk Assessment	Technical + Business + Legal perspectives

---

4. Consensus/Debate Orchestration

Multiple Perspectives: Agents debate and refine positions to reach a higher-quality consensus.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    Debate Architecture                              │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  Round 0: Initial Positions                                         │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐             │
│  │   Agent A    │  │   Agent B    │  │   Agent C    │             │
│  │  (Optimist)  │  │  (Skeptic)   │  │ (Pragmatist) │             │
│  └──────────────┘  └──────────────┘  └──────────────┘             │
│         │                │                  │                      │
│         ▼                ▼                  ▼                      │
│    Position A₀      Position B₀        Position C₀                 │
│                                                                     │
│  Round 1: Read others, refine                                       │
│         │←───────────────┼──────────────────│                      │
│         ▼                ▼                  ▼                      │
│    Position A₁      Position B₁        Position C₁                 │
│         (considers B₀, C₀)                                         │
│                                                                     │
│  Round N: Convergence                                               │
│         │                │                  │                      │
│         ▼                ▼                  ▼                      │
│    ┌─────────────────────────────────────────────┐                 │
│    │              Moderator Agent                 │                 │
│    │   (Synthesize consensus, note dissent)      │                 │
│    └─────────────────────────────────────────────┘                 │
│                          │                                          │
│                          ▼                                          │
│                  Consensus Output                                   │
│                  + Minority Opinions                                │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class DebateOrchestrator {

    private final List<AgentDefinition> debaters;
    private final AgentDefinition moderator;
    private final int maxRounds;
    private final double convergenceThreshold;
    private final MeterRegistry registry;

    public DebateOrchestrator(
            @Qualifier("debaterAgents") List<AgentDefinition> debaters,
            @Qualifier("moderatorAgent") AgentDefinition moderator,
            @Value("${debate.max-rounds:5}") int maxRounds,
            @Value("${debate.convergence-threshold:0.85}") double convergenceThreshold,
            MeterRegistry registry) {
        this.debaters = debaters;
        this.moderator = moderator;
        this.maxRounds = maxRounds;
        this.convergenceThreshold = convergenceThreshold;
        this.registry = registry;
    }

    public DebateResult debate(String topic) {
        log.info("Starting debate on topic with {} participants", debaters.size());

        List<DebateRound> rounds = new ArrayList<>();
        Map<String, String> currentPositions = new HashMap<>();

        // Initial positions
        DebateRound initialRound = collectInitialPositions(topic);
        rounds.add(initialRound);
        initialRound.getPositions().forEach(p ->
            currentPositions.put(p.getAgentName(), p.getPosition()));

        // Debate rounds
        for (int round = 1; round <= maxRounds; round++) {
            log.info("Starting debate round {}/{}", round, maxRounds);

            DebateRound debateRound = executeDebateRound(topic, round, currentPositions);
            rounds.add(debateRound);

            // Update positions
            debateRound.getPositions().forEach(p ->
                currentPositions.put(p.getAgentName(), p.getPosition()));

            // Check for convergence
            double similarity = calculateConvergence(debateRound.getPositions());
            log.info("Round {} convergence score: {}", round, similarity);

            if (similarity >= convergenceThreshold) {
                log.info("Convergence achieved at round {}", round);
                break;
            }
        }

        // Moderator synthesizes
        ConsensusResult consensus = synthesizeConsensus(topic, rounds);

        return DebateResult.builder()
            .topic(topic)
            .rounds(rounds)
            .consensus(consensus)
            .totalRounds(rounds.size())
            .build();
    }

    private DebateRound collectInitialPositions(String topic) {
        List<CompletableFuture<DebatePosition>> futures = debaters.stream()
            .map(agent -> CompletableFuture.supplyAsync(() ->
                getInitialPosition(agent, topic)))
            .collect(Collectors.toList());

        List<DebatePosition> positions = futures.stream()
            .map(CompletableFuture::join)
            .collect(Collectors.toList());

        return DebateRound.builder()
            .roundNumber(0)
            .positions(positions)
            .build();
    }

    private DebatePosition getInitialPosition(AgentDefinition agent, String topic) {
        String position = agent.getChatClient().prompt()
            .system(s -> s.text("""
                You are: {role}
                Perspective: {backstory}

                You are participating in a structured debate. Present your initial
                position on the topic. Be clear, reasoned, and willing to defend
                your viewpoint while remaining open to persuasion.

                Structure your response:
                1. Core Position (1-2 sentences)
                2. Key Arguments (3-5 points)
                3. Evidence/Reasoning
                4. Potential Weaknesses (acknowledge honestly)
                """)
                .param("role", agent.getRole())
                .param("backstory", agent.getBackstory()))
            .user("Topic for debate: " + topic)
            .call()
            .content();

        return DebatePosition.builder()
            .agentName(agent.getName())
            .agentRole(agent.getRole())
            .position(position)
            .build();
    }

    private DebateRound executeDebateRound(String topic, int roundNumber,
                                           Map<String, String> previousPositions) {
        List<CompletableFuture<DebatePosition>> futures = debaters.stream()
            .map(agent -> CompletableFuture.supplyAsync(() ->
                refinePosition(agent, topic, roundNumber, previousPositions)))
            .collect(Collectors.toList());

        List<DebatePosition> positions = futures.stream()
            .map(CompletableFuture::join)
            .collect(Collectors.toList());

        return DebateRound.builder()
            .roundNumber(roundNumber)
            .positions(positions)
            .build();
    }

    private DebatePosition refinePosition(AgentDefinition agent, String topic,
                                          int roundNumber,
                                          Map<String, String> otherPositions) {
        String othersContext = otherPositions.entrySet().stream()
            .filter(e -> !e.getKey().equals(agent.getName()))
            .map(e -> String.format("**%s's position:**\n%s", e.getKey(), e.getValue()))
            .collect(Collectors.joining("\n\n---\n\n"));

        String refinedPosition = agent.getChatClient().prompt()
            .system(s -> s.text("""
                You are: {role}
                Perspective: {backstory}

                This is round {round} of a structured debate.

                Instructions:
                1. Consider the other participants' arguments carefully
                2. Acknowledge strong points from others
                3. Address criticisms of your position
                4. Refine your position based on the discussion
                5. You may change your view if convinced - intellectual honesty is valued

                Structure your response:
                1. Response to Others (engage with specific arguments)
                2. Refined Position (may be unchanged, modified, or reversed)
                3. New Arguments/Evidence
                4. Areas of Agreement (with other debaters)
                5. Remaining Disagreements
                """)
                .param("role", agent.getRole())
                .param("backstory", agent.getBackstory())
                .param("round", roundNumber))
            .user(u -> u.text("""
                Topic: {topic}

                Your previous position:
                {myPosition}

                Other participants' positions:
                {others}

                Provide your refined position for this round.
                """)
                .param("topic", topic)
                .param("myPosition", otherPositions.get(agent.getName()))
                .param("others", othersContext))
            .call()
            .content();

        return DebatePosition.builder()
            .agentName(agent.getName())
            .agentRole(agent.getRole())
            .position(refinedPosition)
            .build();
    }

    private double calculateConvergence(List<DebatePosition> positions) {
        // Use embedding similarity to measure position convergence
        // Simplified: check for explicit agreement markers
        long agreementCount = positions.stream()
            .filter(p -> p.getPosition().toLowerCase().contains("i agree") ||
                        p.getPosition().toLowerCase().contains("consensus") ||
                        p.getPosition().toLowerCase().contains("common ground"))
            .count();

        return (double) agreementCount / positions.size();
    }

    private ConsensusResult synthesizeConsensus(String topic, List<DebateRound> rounds) {
        DebateRound finalRound = rounds.get(rounds.size() - 1);
        String finalPositions = finalRound.getPositions().stream()
            .map(p -> String.format("**%s (%s):**\n%s",
                p.getAgentName(), p.getAgentRole(), p.getPosition()))
            .collect(Collectors.joining("\n\n---\n\n"));

        return moderator.getChatClient().prompt()
            .system("""
                You are a skilled debate moderator and synthesizer.

                After observing a multi-round debate, your task is to:
                1. Identify points of consensus (agreed by most/all)
                2. Highlight valuable insights from each perspective
                3. Note remaining points of disagreement
                4. Provide a balanced synthesis that captures the debate's wisdom
                5. Offer your assessment of the strongest arguments

                Be fair to all perspectives while being honest about argument quality.
                """)
            .user(u -> u.text("""
                Topic: {topic}

                Number of debate rounds: {rounds}

                Final positions from all participants:
                {positions}

                Synthesize the debate into a comprehensive consensus document.
                """)
                .param("topic", topic)
                .param("rounds", rounds.size())
                .param("positions", finalPositions))
            .call()
            .entity(ConsensusResult.class);
    }
}

@Data
@Builder
public class ConsensusResult {
    private String executiveSummary;
    private List<String> pointsOfAgreement;
    private List<String> pointsOfDisagreement;
    private List<String> keyInsights;
    private String recommendedAction;
    private double confidenceLevel;
}

Best For

Use Case	Example
Decision Making	Investment decisions, hiring
Risk Assessment	Project go/no-go decisions
Policy Development	Company guidelines, standards
Complex Analysis	Multiple valid interpretations

---

5. ReAct: Reasoning and Acting

The Thinking Actor: Interleaves reasoning (thinking about what to do) with acting (taking actions and observing results).

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                       ReAct Loop                                    │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  Task ──▶ ┌─────────────────────────────────────────────────────┐  │
│           │                                                     │  │
│           │  ┌─────────┐    ┌─────────┐    ┌─────────────────┐ │  │
│           │  │ Thought │ ─▶ │ Action  │ ─▶ │   Observation   │ │  │
│           │  │ (Plan)  │    │(Execute)│    │    (Result)     │ │  │
│           │  └─────────┘    └─────────┘    └─────────────────┘ │  │
│           │       ▲                              │              │  │
│           │       │                              │              │  │
│           │       └──────────────────────────────┘              │  │
│           │                                                     │  │
│           │  Repeat until task complete or max iterations      │  │
│           └─────────────────────────────────────────────────────┘  │
│                                      │                              │
│                                      ▼                              │
│                               Final Answer                          │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class ReActAgent {

    private final ChatClient chatClient;
    private final Map<String, ToolExecutor> tools;
    private final int maxIterations;
    private final MeterRegistry registry;

    public ReActAgent(
            ChatClient.Builder builder,
            Map<String, ToolExecutor> tools,
            @Value("${react.max-iterations:10}") int maxIterations,
            MeterRegistry registry) {
        this.chatClient = builder.build();
        this.tools = tools;
        this.maxIterations = maxIterations;
        this.registry = registry;
    }

    public ReActResult execute(String task) {
        log.info("Starting ReAct agent for task: {}",
            task.substring(0, Math.min(100, task.length())));

        List<ReActStep> trajectory = new ArrayList<>();
        StringBuilder scratchpad = new StringBuilder();

        String availableTools = tools.keySet().stream()
            .map(name -> String.format("- %s: %s", name, tools.get(name).getDescription()))
            .collect(Collectors.joining("\n"));

        for (int i = 0; i < maxIterations; i++) {
            log.info("ReAct iteration {}/{}", i + 1, maxIterations);

            // Get next thought and action
            ReActResponse response = getNextStep(task, scratchpad.toString(), availableTools);

            if (response.isFinished()) {
                log.info("ReAct agent completed in {} iterations", i + 1);
                return ReActResult.builder()
                    .task(task)
                    .answer(response.getFinalAnswer())
                    .trajectory(trajectory)
                    .iterations(i + 1)
                    .success(true)
                    .build();
            }

            // Execute action
            String observation;
            try {
                observation = executeAction(response.getAction(), response.getActionInput());
            } catch (Exception e) {
                observation = "Error: " + e.getMessage();
            }

            // Record step
            ReActStep step = ReActStep.builder()
                .iteration(i + 1)
                .thought(response.getThought())
                .action(response.getAction())
                .actionInput(response.getActionInput())
                .observation(observation)
                .build();
            trajectory.add(step);

            // Update scratchpad
            scratchpad.append(String.format("""

                Thought %d: %s
                Action %d: %s[%s]
                Observation %d: %s
                """,
                i + 1, response.getThought(),
                i + 1, response.getAction(), response.getActionInput(),
                i + 1, observation
            ));
        }

        log.warn("ReAct agent reached max iterations without completing");

        // Force final answer
        String forcedAnswer = getForcedAnswer(task, scratchpad.toString());

        return ReActResult.builder()
            .task(task)
            .answer(forcedAnswer)
            .trajectory(trajectory)
            .iterations(maxIterations)
            .success(false)
            .reason("Max iterations reached")
            .build();
    }

    private ReActResponse getNextStep(String task, String scratchpad, String tools) {
        String response = chatClient.prompt()
            .system(s -> s.text("""
                You are a reasoning agent that solves tasks by thinking step-by-step
                and using tools to gather information.

                Available Tools:
                {tools}

                Response Format:
                Thought: [Your reasoning about what to do next]
                Action: [Tool name to use, or "Finish" if done]
                Action Input: [Input for the tool, or final answer if Action is Finish]

                Important:
                - Think carefully before each action
                - Use observations to inform next steps
                - Finish when you have enough information to answer
                - Be concise but thorough
                """)
                .param("tools", tools))
            .user(u -> u.text("""
                Task: {task}

                Previous steps:
                {scratchpad}

                What is your next thought and action?
                """)
                .param("task", task)
                .param("scratchpad", scratchpad.isEmpty() ? "None yet" : scratchpad))
            .call()
            .content();

        return parseReActResponse(response);
    }

    private ReActResponse parseReActResponse(String response) {
        // Parse the structured response
        String thought = extractField(response, "Thought:");
        String action = extractField(response, "Action:");
        String actionInput = extractField(response, "Action Input:");

        boolean finished = action.equalsIgnoreCase("Finish") ||
                          action.equalsIgnoreCase("Final Answer");

        return ReActResponse.builder()
            .thought(thought)
            .action(action)
            .actionInput(actionInput)
            .finished(finished)
            .finalAnswer(finished ? actionInput : null)
            .build();
    }

    private String executeAction(String action, String input) {
        ToolExecutor executor = tools.get(action.toLowerCase());

        if (executor == null) {
            return String.format("Unknown tool: %s. Available: %s",
                action, String.join(", ", tools.keySet()));
        }

        Timer.Sample timer = Timer.start(registry);

        try {
            String result = executor.execute(input);
            timer.stop(registry.timer("react.tool", "tool", action, "status", "success"));
            return result;
        } catch (Exception e) {
            timer.stop(registry.timer("react.tool", "tool", action, "status", "error"));
            throw e;
        }
    }

    private String extractField(String response, String fieldName) {
        int start = response.indexOf(fieldName);
        if (start == -1) return "";

        start += fieldName.length();
        int end = response.indexOf("\n", start);
        if (end == -1) end = response.length();

        return response.substring(start, end).trim();
    }

    private String getForcedAnswer(String task, String scratchpad) {
        return chatClient.prompt()
            .system("""
                Based on the reasoning trajectory, provide the best possible answer
                to the original task. Acknowledge any limitations or uncertainties.
                """)
            .user(u -> u.text("""
                Task: {task}

                Reasoning trajectory:
                {scratchpad}

                Provide your best answer.
                """)
                .param("task", task)
                .param("scratchpad", scratchpad))
            .call()
            .content();
    }
}

@FunctionalInterface
public interface ToolExecutor {
    String execute(String input);

    default String getDescription() {
        return "A tool";
    }
}

// Tool implementations
@Component("search")
public class WebSearchTool implements ToolExecutor {

    private final WebSearchService searchService;

    @Override
    public String execute(String query) {
        return searchService.search(query).stream()
            .limit(5)
            .map(r -> String.format("- %s: %s", r.getTitle(), r.getSnippet()))
            .collect(Collectors.joining("\n"));
    }

    @Override
    public String getDescription() {
        return "Search the web for information. Input: search query";
    }
}

@Component("calculator")
public class CalculatorTool implements ToolExecutor {

    @Override
    public String execute(String expression) {
        try {
            // Safe expression evaluation
            ScriptEngine engine = new ScriptEngineManager().getEngineByName("JavaScript");
            return String.valueOf(engine.eval(expression));
        } catch (Exception e) {
            return "Error evaluating: " + e.getMessage();
        }
    }

    @Override
    public String getDescription() {
        return "Calculate mathematical expressions. Input: math expression";
    }
}

---

6. Plan-and-Execute Pattern

The Strategic Planner: First creates a complete plan, then executes each step, re-planning as needed.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    Plan-and-Execute Architecture                    │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  Task ──▶ ┌─────────────────────────────────────────────────────┐  │
│           │              Planning Phase                         │  │
│           │  ┌─────────────────────────────────────────────────┐│  │
│           │  │ 1. Analyze task requirements                    ││  │
│           │  │ 2. Identify required steps                      ││  │
│           │  │ 3. Order steps by dependencies                  ││  │
│           │  │ 4. Assign resources/tools per step              ││  │
│           │  └─────────────────────────────────────────────────┘│  │
│           └────────────────────────┬────────────────────────────┘  │
│                                    │                                │
│                                    ▼                                │
│           ┌─────────────────────────────────────────────────────┐  │
│           │              Execution Phase                        │  │
│           │  ┌──────────────────────────────────────────────┐  │  │
│           │  │ For each step in plan:                       │  │  │
│           │  │   1. Execute step                            │  │  │
│           │  │   2. Observe result                          │  │  │
│           │  │   3. Check if re-planning needed             │  │  │
│           │  │   4. Update remaining plan if necessary      │  │  │
│           │  └──────────────────────────────────────────────┘  │  │
│           └────────────────────────┬────────────────────────────┘  │
│                                    │                                │
│                                    ▼                                │
│                              Final Result                           │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Implementation

@Service
@Slf4j
public class PlanAndExecuteAgent {

    private final ChatClient planner;
    private final ChatClient executor;
    private final ChatClient replanner;
    private final Map<String, ToolExecutor> tools;
    private final MeterRegistry registry;

    public PlanExecuteResult execute(String task) {
        log.info("Starting Plan-and-Execute for task");

        // Phase 1: Create initial plan
        ExecutionPlan plan = createPlan(task);
        log.info("Created plan with {} steps", plan.getSteps().size());

        List<StepResult> results = new ArrayList<>();

        // Phase 2: Execute each step
        for (int i = 0; i < plan.getSteps().size(); i++) {
            PlanStep step = plan.getSteps().get(i);
            log.info("Executing step {}/{}: {}", i + 1, plan.getSteps().size(), step.getDescription());

            // Execute step
            StepResult result = executeStep(step, results);
            results.add(result);

            // Check if re-planning is needed
            if (!result.isSuccess() || result.isRequiresReplan()) {
                log.info("Re-planning required after step {}", i + 1);

                ExecutionPlan newPlan = replan(task, plan, results, i);

                if (newPlan != null && !newPlan.getSteps().isEmpty()) {
                    plan = newPlan;
                    // Don't increment i - will re-evaluate from current position
                    continue;
                } else if (!result.isSuccess()) {
                    return PlanExecuteResult.builder()
                        .task(task)
                        .originalPlan(plan)
                        .stepResults(results)
                        .success(false)
                        .error("Failed at step " + (i + 1) + ": " + result.getError())
                        .build();
                }
            }
        }

        // Phase 3: Synthesize final result
        String finalResult = synthesizeResult(task, results);

        return PlanExecuteResult.builder()
            .task(task)
            .originalPlan(plan)
            .stepResults(results)
            .finalResult(finalResult)
            .success(true)
            .build();
    }

    private ExecutionPlan createPlan(String task) {
        String toolDescriptions = tools.entrySet().stream()
            .map(e -> String.format("- %s: %s", e.getKey(), e.getValue().getDescription()))
            .collect(Collectors.joining("\n"));

        return planner.prompt()
            .system(s -> s.text("""
                You are a strategic planner. Create a detailed execution plan for the task.

                Available tools:
                {tools}

                Plan requirements:
                1. Break task into atomic, executable steps
                2. Each step should use one tool or be a synthesis step
                3. Order steps by logical dependencies
                4. Include validation/verification steps
                5. Consider failure scenarios

                Output a JSON plan with:
                - steps: array of {description, tool, expectedOutput, dependsOn[]}
                - successCriteria: how to know the task is complete
                - riskFactors: potential issues to watch for
                """)
                .param("tools", toolDescriptions))
            .user(task)
            .call()
            .entity(ExecutionPlan.class);
    }

    private StepResult executeStep(PlanStep step, List<StepResult> previousResults) {
        Timer.Sample timer = Timer.start(registry);

        try {
            // Gather context from previous steps
            String context = previousResults.stream()
                .filter(r -> step.getDependsOn().contains(r.getStepId()))
                .map(r -> String.format("From %s: %s", r.getStepId(), r.getOutput()))
                .collect(Collectors.joining("\n"));

            // Execute the step
            String output;
            if (step.getTool() != null && tools.containsKey(step.getTool())) {
                // Tool execution
                String toolInput = prepareToolInput(step, context);
                output = tools.get(step.getTool()).execute(toolInput);
            } else {
                // LLM execution (synthesis, analysis, etc.)
                output = executor.prompt()
                    .system("""
                        You are executing a step in a larger plan.
                        Use the provided context and complete the step thoroughly.
                        """)
                    .user(u -> u.text("""
                        Step: {description}

                        Context from previous steps:
                        {context}

                        Expected output: {expected}

                        Execute this step and provide the result.
                        """)
                        .param("description", step.getDescription())
                        .param("context", context.isEmpty() ? "None" : context)
                        .param("expected", step.getExpectedOutput()))
                    .call()
                    .content();
            }

            // Validate output
            boolean meetsExpectation = validateOutput(output, step.getExpectedOutput());

            timer.stop(registry.timer("planexec.step", "status", "success"));

            return StepResult.builder()
                .stepId(step.getId())
                .description(step.getDescription())
                .output(output)
                .success(true)
                .meetsExpectation(meetsExpectation)
                .requiresReplan(!meetsExpectation)
                .build();

        } catch (Exception e) {
            timer.stop(registry.timer("planexec.step", "status", "error"));

            return StepResult.builder()
                .stepId(step.getId())
                .description(step.getDescription())
                .success(false)
                .error(e.getMessage())
                .requiresReplan(true)
                .build();
        }
    }

    private ExecutionPlan replan(String task, ExecutionPlan currentPlan,
                                  List<StepResult> results, int failedStepIndex) {
        String completedSteps = results.stream()
            .map(r -> String.format("- %s: %s (success: %s)",
                r.getDescription(), r.getOutput(), r.isSuccess()))
            .collect(Collectors.joining("\n"));

        String remainingSteps = currentPlan.getSteps().subList(
            failedStepIndex, currentPlan.getSteps().size()).stream()
            .map(s -> "- " + s.getDescription())
            .collect(Collectors.joining("\n"));

        return replanner.prompt()
            .system("""
                You are re-planning after a step failed or produced unexpected results.

                Consider:
                1. What went wrong?
                2. Can we recover or need alternative approach?
                3. Are remaining steps still valid?
                4. What adjustments are needed?
                """)
            .user(u -> u.text("""
                Original task: {task}

                Completed steps:
                {completed}

                Remaining planned steps:
                {remaining}

                Issue: Step failed or produced unexpected output

                Create a revised plan for the remaining work.
                Return null if the task cannot be completed.
                """)
                .param("task", task)
                .param("completed", completedSteps)
                .param("remaining", remainingSteps))
            .call()
            .entity(ExecutionPlan.class);
    }

    private String synthesizeResult(String task, List<StepResult> results) {
        String allResults = results.stream()
            .map(r -> String.format("**%s:**\n%s", r.getDescription(), r.getOutput()))
            .collect(Collectors.joining("\n\n"));

        return executor.prompt()
            .system("""
                Synthesize the results of all completed steps into a final,
                coherent response that addresses the original task.
                """)
            .user(u -> u.text("""
                Original task: {task}

                Completed steps and results:
                {results}

                Provide the final answer.
                """)
                .param("task", task)
                .param("results", allResults))
            .call()
            .content();
    }

    private String prepareToolInput(PlanStep step, String context) {
        return executor.prompt()
            .system("Prepare the input for the tool based on the step description and context.")
            .user(u -> u.text("""
                Step: {description}
                Tool: {tool}
                Context: {context}

                What should be the input to the tool?
                """)
                .param("description", step.getDescription())
                .param("tool", step.getTool())
                .param("context", context))
            .call()
            .content();
    }

    private boolean validateOutput(String output, String expectedOutput) {
        // Simple validation - could be enhanced with LLM-based validation
        return output != null && !output.isEmpty() &&
               !output.toLowerCase().contains("error") &&
               !output.toLowerCase().contains("failed");
    }
}

@Data
@Builder
public class ExecutionPlan {
    private List<PlanStep> steps;
    private String successCriteria;
    private List<String> riskFactors;
}

@Data
@Builder
public class PlanStep {
    private String id;
    private String description;
    private String tool;
    private String expectedOutput;
    private List<String> dependsOn;
}

---

7. Tool Integration with MCP

Model Context Protocol (MCP): Standard protocol for connecting AI agents to external tools and data sources.

MCP Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    MCP Integration Architecture                     │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │                      AI Agent (Host)                        │   │
│  │  ┌─────────────────────────────────────────────────────────┐│   │
│  │  │                   MCP Client                             ││   │
│  │  │  - Discovers available servers                          ││   │
│  │  │  - Routes tool calls                                    ││   │
│  │  │  - Manages connections                                  ││   │
│  │  └─────────────────────────────────────────────────────────┘│   │
│  └────────────────────────┬────────────────────────────────────┘   │
│                           │                                         │
│         ┌─────────────────┼─────────────────┐                      │
│         │                 │                 │                       │
│         ▼                 ▼                 ▼                       │
│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐              │
│  │ MCP Server  │   │ MCP Server  │   │ MCP Server  │              │
│  │  (GitHub)   │   │ (Database)  │   │  (Custom)   │              │
│  │             │   │             │   │             │              │
│  │ - Tools     │   │ - Tools     │   │ - Tools     │              │
│  │ - Resources │   │ - Resources │   │ - Resources │              │
│  │ - Prompts   │   │ - Prompts   │   │ - Prompts   │              │
│  └─────────────┘   └─────────────┘   └─────────────┘              │
│         │                 │                 │                       │
│         ▼                 ▼                 ▼                       │
│     External          Database          Custom                      │
│      APIs                                Services                    │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI MCP Integration

@Configuration
public class McpConfiguration {

    @Bean
    public McpClient mcpClient(
            @Value("${mcp.servers}") List<String> serverUrls) {
        return McpClient.builder()
            .servers(serverUrls.stream()
                .map(url -> McpServerConfig.builder()
                    .url(url)
                    .transport(McpTransport.STDIO)
                    .build())
                .collect(Collectors.toList()))
            .build();
    }
}

@Service
@Slf4j
public class McpToolRegistry {

    private final McpClient mcpClient;
    private final Map<String, McpTool> registeredTools = new ConcurrentHashMap<>();

    @PostConstruct
    public void discoverTools() {
        log.info("Discovering MCP tools from connected servers");

        mcpClient.listTools().forEach(tool -> {
            registeredTools.put(tool.getName(), tool);
            log.info("Registered MCP tool: {} - {}",
                tool.getName(), tool.getDescription());
        });

        log.info("Discovered {} MCP tools", registeredTools.size());
    }

    public List<FunctionCallback> asFunctionCallbacks() {
        return registeredTools.values().stream()
            .map(this::toFunctionCallback)
            .collect(Collectors.toList());
    }

    private FunctionCallback toFunctionCallback(McpTool tool) {
        return FunctionCallback.builder()
            .name(tool.getName())
            .description(tool.getDescription())
            .inputSchema(tool.getInputSchema())
            .executor(input -> executeMcpTool(tool.getName(), input))
            .build();
    }

    public String executeMcpTool(String toolName, Map<String, Object> input) {
        McpTool tool = registeredTools.get(toolName);
        if (tool == null) {
            throw new IllegalArgumentException("Unknown MCP tool: " + toolName);
        }

        try {
            McpToolResult result = mcpClient.callTool(
                tool.getServerName(),
                toolName,
                input
            );

            return result.getContent().stream()
                .map(McpContent::getText)
                .collect(Collectors.joining("\n"));

        } catch (Exception e) {
            log.error("MCP tool execution failed: {}", e.getMessage());
            throw new RuntimeException("Tool execution failed: " + e.getMessage(), e);
        }
    }
}

// Agent with MCP Tools
@Service
public class McpEnabledAgent {

    private final ChatClient chatClient;
    private final McpToolRegistry toolRegistry;

    public McpEnabledAgent(ChatClient.Builder builder, McpToolRegistry toolRegistry) {
        this.toolRegistry = toolRegistry;
        this.chatClient = builder
            .defaultFunctions(toolRegistry.asFunctionCallbacks())
            .build();
    }

    public String execute(String task) {
        return chatClient.prompt()
            .system("""
                You are an AI assistant with access to various tools.
                Use tools when they can help accomplish the task.
                Think step-by-step and use tools as needed.
                """)
            .user(task)
            .call()
            .content();
    }
}

Custom MCP Server Implementation

@SpringBootApplication
public class CustomMcpServer {

    public static void main(String[] args) {
        SpringApplication.run(CustomMcpServer.class, args);
    }
}

@Configuration
public class McpServerConfiguration {

    @Bean
    public McpServer mcpServer(List<McpToolProvider> toolProviders) {
        return McpServer.builder()
            .name("custom-tools")
            .version("1.0.0")
            .tools(toolProviders.stream()
                .flatMap(p -> p.getTools().stream())
                .collect(Collectors.toList()))
            .build();
    }
}

@Component
public class DatabaseToolProvider implements McpToolProvider {

    private final JdbcTemplate jdbcTemplate;

    @Override
    public List<McpToolDefinition> getTools() {
        return List.of(
            McpToolDefinition.builder()
                .name("query_database")
                .description("Execute a read-only SQL query against the database")
                .inputSchema(JsonSchema.builder()
                    .type("object")
                    .property("query", JsonSchema.string()
                        .description("SQL SELECT query to execute"))
                    .required("query")
                    .build())
                .handler(this::executeQuery)
                .build(),

            McpToolDefinition.builder()
                .name("list_tables")
                .description("List all tables in the database")
                .inputSchema(JsonSchema.builder()
                    .type("object")
                    .build())
                .handler(this::listTables)
                .build()
        );
    }

    private McpToolResult executeQuery(Map<String, Object> input) {
        String query = (String) input.get("query");

        // Security: Only allow SELECT queries
        if (!query.trim().toUpperCase().startsWith("SELECT")) {
            return McpToolResult.error("Only SELECT queries are allowed");
        }

        try {
            List<Map<String, Object>> results = jdbcTemplate.queryForList(query);
            String jsonResult = new ObjectMapper().writeValueAsString(results);
            return McpToolResult.success(jsonResult);
        } catch (Exception e) {
            return McpToolResult.error("Query failed: " + e.getMessage());
        }
    }

    private McpToolResult listTables(Map<String, Object> input) {
        List<String> tables = jdbcTemplate.queryForList(
            "SELECT table_name FROM information_schema.tables WHERE table_schema = 'public'",
            String.class
        );
        return McpToolResult.success(String.join(", ", tables));
    }
}

---

8. Memory and State Management

Memory Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                    Agent Memory Architecture                        │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │                  Short-Term Memory                          │   │
│  │  - Current conversation                                     │   │
│  │  - Working context                                          │   │
│  │  - Recent tool results                                      │   │
│  │  Capacity: ~100K tokens, Duration: Session                  │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                           │                                         │
│                           ▼                                         │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │                 Episodic Memory                             │   │
│  │  - Past interactions summaries                              │   │
│  │  - Successful task patterns                                 │   │
│  │  - User preferences learned                                 │   │
│  │  Storage: Vector DB, Duration: Long-term                    │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                           │                                         │
│                           ▼                                         │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │                 Semantic Memory                             │   │
│  │  - Domain knowledge                                         │   │
│  │  - Facts and relationships                                  │   │
│  │  - Procedural knowledge                                     │   │
│  │  Storage: Knowledge Graph, Duration: Permanent              │   │
│  └─────────────────────────────────────────────────────────────┘   │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Spring AI Memory Implementation

@Service
@Slf4j
public class AgentMemoryService {

    private final VectorStore vectorStore;
    private final ChatClient chatClient;
    private final Map<String, ConversationMemory> shortTermMemory;

    public AgentMemoryService(VectorStore vectorStore, ChatClient.Builder builder) {
        this.vectorStore = vectorStore;
        this.chatClient = builder.build();
        this.shortTermMemory = new ConcurrentHashMap<>();
    }

    // Short-term memory management
    public void addToShortTermMemory(String sessionId, Message message) {
        shortTermMemory.computeIfAbsent(sessionId, k -> new ConversationMemory())
            .addMessage(message);
    }

    public List<Message> getConversationHistory(String sessionId, int maxMessages) {
        ConversationMemory memory = shortTermMemory.get(sessionId);
        if (memory == null) return List.of();

        List<Message> messages = memory.getMessages();
        int start = Math.max(0, messages.size() - maxMessages);
        return messages.subList(start, messages.size());
    }

    // Episodic memory - store and retrieve past experiences
    public void storeEpisode(String sessionId, Episode episode) {
        // Summarize the episode
        String summary = summarizeEpisode(episode);

        // Create embedding and store
        Document doc = Document.builder()
            .content(summary)
            .metadata(Map.of(
                "sessionId", sessionId,
                "timestamp", episode.getTimestamp().toString(),
                "type", "episode",
                "outcome", episode.getOutcome(),
                "task", episode.getTask()
            ))
            .build();

        vectorStore.add(List.of(doc));
        log.info("Stored episode for session {}: {}", sessionId,
            episode.getTask().substring(0, Math.min(50, episode.getTask().length())));
    }

    public List<Episode> retrieveRelevantEpisodes(String query, int limit) {
        List<Document> results = vectorStore.similaritySearch(
            SearchRequest.query(query)
                .withTopK(limit)
                .withFilterExpression("type == 'episode'")
        );

        return results.stream()
            .map(this::documentToEpisode)
            .collect(Collectors.toList());
    }

    // Semantic memory - domain knowledge
    public void storeKnowledge(String domain, String knowledge) {
        Document doc = Document.builder()
            .content(knowledge)
            .metadata(Map.of(
                "type", "knowledge",
                "domain", domain,
                "timestamp", Instant.now().toString()
            ))
            .build();

        vectorStore.add(List.of(doc));
    }

    public List<String> retrieveKnowledge(String query, String domain, int limit) {
        FilterExpressionBuilder builder = new FilterExpressionBuilder();

        List<Document> results = vectorStore.similaritySearch(
            SearchRequest.query(query)
                .withTopK(limit)
                .withFilterExpression(
                    builder.and(
                        builder.eq("type", "knowledge"),
                        builder.eq("domain", domain)
                    ).build()
                )
        );

        return results.stream()
            .map(Document::getContent)
            .collect(Collectors.toList());
    }

    // Memory consolidation - summarize and compress
    public void consolidateMemory(String sessionId) {
        ConversationMemory memory = shortTermMemory.get(sessionId);
        if (memory == null || memory.getMessages().size() < 10) return;

        // Summarize conversation so far
        String summary = chatClient.prompt()
            .system("""
                Summarize the following conversation, capturing:
                1. Main topics discussed
                2. Key decisions made
                3. Important facts learned
                4. User preferences observed
                5. Outcomes of any tasks

                Be concise but comprehensive.
                """)
            .user(formatMessages(memory.getMessages()))
            .call()
            .content();

        // Store as episodic memory
        Episode episode = Episode.builder()
            .sessionId(sessionId)
            .task("Conversation summary")
            .summary(summary)
            .outcome("completed")
            .timestamp(Instant.now())
            .build();

        storeEpisode(sessionId, episode);

        // Trim short-term memory
        memory.trimToSize(5); // Keep only last 5 messages

        log.info("Consolidated memory for session {}", sessionId);
    }

    private String summarizeEpisode(Episode episode) {
        return chatClient.prompt()
            .system("Create a concise summary of this agent episode for future reference.")
            .user(u -> u.text("""
                Task: {task}
                Actions taken: {actions}
                Outcome: {outcome}
                Learnings: {learnings}
                """)
                .param("task", episode.getTask())
                .param("actions", String.join(", ", episode.getActions()))
                .param("outcome", episode.getOutcome())
                .param("learnings", String.join(", ", episode.getLearnings())))
            .call()
            .content();
    }

    private String formatMessages(List<Message> messages) {
        return messages.stream()
            .map(m -> m.getRole() + ": " + m.getContent())
            .collect(Collectors.joining("\n"));
    }

    private Episode documentToEpisode(Document doc) {
        return Episode.builder()
            .sessionId((String) doc.getMetadata().get("sessionId"))
            .task((String) doc.getMetadata().get("task"))
            .summary(doc.getContent())
            .outcome((String) doc.getMetadata().get("outcome"))
            .timestamp(Instant.parse((String) doc.getMetadata().get("timestamp")))
            .build();
    }
}

@Data
@Builder
public class Episode {
    private String sessionId;
    private String task;
    private List<String> actions;
    private String outcome;
    private String summary;
    private List<String> learnings;
    private Instant timestamp;
}

public class ConversationMemory {
    private final List<Message> messages = new ArrayList<>();
    private int maxSize = 100;

    public synchronized void addMessage(Message message) {
        messages.add(message);
        if (messages.size() > maxSize) {
            messages.remove(0);
        }
    }

    public synchronized List<Message> getMessages() {
        return new ArrayList<>(messages);
    }

    public synchronized void trimToSize(int size) {
        while (messages.size() > size) {
            messages.remove(0);
        }
    }
}

---

9. Error Handling and Recovery

Resilience Patterns

@Service
@Slf4j
public class ResilientAgentExecutor {

    private final RetryTemplate retryTemplate;
    private final CircuitBreaker circuitBreaker;
    private final MeterRegistry registry;

    public ResilientAgentExecutor(MeterRegistry registry) {
        this.registry = registry;

        this.retryTemplate = RetryTemplate.builder()
            .maxAttempts(3)
            .exponentialBackoff(1000, 2, 10000)
            .retryOn(TransientException.class)
            .build();

        this.circuitBreaker = CircuitBreaker.of("agent-executor",
            CircuitBreakerConfig.custom()
                .failureRateThreshold(50)
                .waitDurationInOpenState(Duration.ofSeconds(30))
                .slidingWindowSize(10)
                .build());
    }

    public <T> T executeWithResilience(String operationName,
                                        Supplier<T> operation,
                                        Supplier<T> fallback) {
        Timer.Sample timer = Timer.start(registry);

        try {
            T result = circuitBreaker.executeSupplier(() ->
                retryTemplate.execute(context -> {
                    if (context.getRetryCount() > 0) {
                        log.warn("Retry attempt {} for operation: {}",
                            context.getRetryCount(), operationName);
                        registry.counter("agent.retry",
                            "operation", operationName).increment();
                    }
                    return operation.get();
                })
            );

            timer.stop(registry.timer("agent.operation",
                "operation", operationName, "status", "success"));
            return result;

        } catch (Exception e) {
            timer.stop(registry.timer("agent.operation",
                "operation", operationName, "status", "error"));

            log.error("Operation {} failed after retries: {}", operationName, e.getMessage());

            if (fallback != null) {
                log.info("Executing fallback for operation: {}", operationName);
                registry.counter("agent.fallback", "operation", operationName).increment();
                return fallback.get();
            }

            throw new AgentExecutionException("Operation failed: " + operationName, e);
        }
    }

    public <T> CompletableFuture<T> executeWithTimeout(
            Supplier<T> operation,
            Duration timeout,
            T defaultValue) {
        return CompletableFuture.supplyAsync(operation)
            .completeOnTimeout(defaultValue, timeout.toMillis(), TimeUnit.MILLISECONDS)
            .exceptionally(e -> {
                log.error("Async operation failed: {}", e.getMessage());
                return defaultValue;
            });
    }
}

// Self-healing agent with error recovery
@Service
@Slf4j
public class SelfHealingAgent {

    private final ChatClient chatClient;
    private final ResilientAgentExecutor executor;
    private final Map<String, RecoveryStrategy> recoveryStrategies;

    public AgentResult executeWithRecovery(String task) {
        List<AgentAttempt> attempts = new ArrayList<>();

        for (int attempt = 0; attempt < 3; attempt++) {
            try {
                String result = executor.executeWithResilience(
                    "agent-task",
                    () -> executeTask(task, attempts),
                    null
                );

                return AgentResult.builder()
                    .success(true)
                    .output(result)
                    .attempts(attempts)
                    .build();

            } catch (AgentExecutionException e) {
                AgentAttempt attemptRecord = AgentAttempt.builder()
                    .attemptNumber(attempt + 1)
                    .error(e.getMessage())
                    .timestamp(Instant.now())
                    .build();
                attempts.add(attemptRecord);

                // Analyze error and determine recovery strategy
                RecoveryAction action = analyzeAndRecover(task, e, attempts);

                if (action.shouldRetry()) {
                    task = action.getModifiedTask() != null ?
                        action.getModifiedTask() : task;
                    log.info("Retrying with modified approach: {}", action.getStrategy());
                } else {
                    break;
                }
            }
        }

        // All attempts failed
        return AgentResult.builder()
            .success(false)
            .error("Task failed after all recovery attempts")
            .attempts(attempts)
            .build();
    }

    private String executeTask(String task, List<AgentAttempt> previousAttempts) {
        String context = previousAttempts.isEmpty() ? "" :
            "Previous attempts failed. Avoid these approaches: " +
            previousAttempts.stream()
                .map(a -> a.getError())
                .collect(Collectors.joining("; "));

        return chatClient.prompt()
            .system(s -> s.text("""
                You are an AI agent executing a task.
                {context}

                If you encounter errors:
                1. Explain what went wrong
                2. Suggest alternative approaches
                3. Try a different strategy if current one fails
                """)
                .param("context", context))
            .user(task)
            .call()
            .content();
    }

    private RecoveryAction analyzeAndRecover(String task, Exception error,
                                              List<AgentAttempt> attempts) {
        // Analyze error pattern
        String errorAnalysis = chatClient.prompt()
            .system("""
                Analyze this error and suggest a recovery strategy.

                Possible strategies:
                - SIMPLIFY: Break down the task into simpler steps
                - ALTERNATIVE: Use a different approach entirely
                - RESOURCE: The error is due to external resource, retry later
                - CLARIFY: The task is ambiguous, need clarification
                - ABORT: The task cannot be completed

                Return JSON: {strategy, reason, modifiedTask}
                """)
            .user(u -> u.text("""
                Task: {task}
                Error: {error}
                Previous attempts: {attempts}
                """)
                .param("task", task)
                .param("error", error.getMessage())
                .param("attempts", attempts.size()))
            .call()
            .content();

        // Parse and return recovery action
        return parseRecoveryAction(errorAnalysis);
    }

    private RecoveryAction parseRecoveryAction(String analysis) {
        try {
            return new ObjectMapper().readValue(analysis, RecoveryAction.class);
        } catch (Exception e) {
            return RecoveryAction.builder()
                .strategy("ABORT")
                .reason("Could not parse recovery strategy")
                .shouldRetry(false)
                .build();
        }
    }
}

@Data
@Builder
public class RecoveryAction {
    private String strategy;
    private String reason;
    private String modifiedTask;
    private boolean shouldRetry;
}

---

10. Production Patterns

Observability and Monitoring

@Configuration
public class AgentObservabilityConfig {

    @Bean
    public MeterRegistry meterRegistry() {
        return new PrometheusMeterRegistry(PrometheusConfig.DEFAULT);
    }

    @Bean
    public TracingConfig tracingConfig() {
        return TracingConfig.builder()
            .serviceName("ai-agent-service")
            .samplingRate(0.1) // 10% sampling in production
            .build();
    }
}

@Aspect
@Component
@Slf4j
public class AgentTracingAspect {

    private final Tracer tracer;
    private final MeterRegistry registry;

    @Around("@annotation(AgentOperation)")
    public Object traceAgentOperation(ProceedingJoinPoint joinPoint) throws Throwable {
        String operationName = joinPoint.getSignature().getName();

        Span span = tracer.nextSpan()
            .name(operationName)
            .tag("agent.operation", operationName)
            .start();

        Timer.Sample timer = Timer.start(registry);

        try (Tracer.SpanInScope ws = tracer.withSpan(span)) {
            Object result = joinPoint.proceed();

            span.tag("status", "success");
            timer.stop(registry.timer("agent.operation",
                "name", operationName, "status", "success"));

            return result;

        } catch (Exception e) {
            span.tag("status", "error")
                .tag("error.message", e.getMessage());
            timer.stop(registry.timer("agent.operation",
                "name", operationName, "status", "error"));

            throw e;

        } finally {
            span.end();
        }
    }
}

// Structured logging for agents
@Service
@Slf4j
public class AgentLogger {

    public void logAgentAction(String agentName, String action,
                               Map<String, Object> context) {
        log.info("agent_action",
            kv("agent_name", agentName),
            kv("action", action),
            kv("context", context),
            kv("timestamp", Instant.now())
        );
    }

    public void logToolExecution(String toolName, String input,
                                  String output, long durationMs) {
        log.info("tool_execution",
            kv("tool_name", toolName),
            kv("input_length", input.length()),
            kv("output_length", output.length()),
            kv("duration_ms", durationMs)
        );
    }

    public void logOrchestrationEvent(String orchestrationType,
                                       String phase,
                                       Map<String, Object> metrics) {
        log.info("orchestration_event",
            kv("type", orchestrationType),
            kv("phase", phase),
            kv("metrics", metrics)
        );
    }

    private StructuredArgument kv(String key, Object value) {
        return StructuredArguments.kv(key, value);
    }
}

Cost Management

@Service
@Slf4j
public class AgentCostManager {

    private final Map<String, ModelPricing> modelPricing;
    private final MeterRegistry registry;
    private final AtomicLong totalCostCents = new AtomicLong(0);

    public AgentCostManager(MeterRegistry registry) {
        this.registry = registry;
        this.modelPricing = Map.of(
            "gpt-4-turbo", new ModelPricing(10.0, 30.0),      // per 1M tokens
            "gpt-4o", new ModelPricing(5.0, 15.0),
            "gpt-4o-mini", new ModelPricing(0.15, 0.6),
            "claude-3-5-sonnet", new ModelPricing(3.0, 15.0),
            "claude-3-5-haiku", new ModelPricing(0.25, 1.25)
        );
    }

    public void trackUsage(String model, int inputTokens, int outputTokens) {
        ModelPricing pricing = modelPricing.getOrDefault(model,
            new ModelPricing(1.0, 3.0));

        double inputCost = (inputTokens / 1_000_000.0) * pricing.inputPricePerMillion();
        double outputCost = (outputTokens / 1_000_000.0) * pricing.outputPricePerMillion();
        double totalCost = inputCost + outputCost;

        long costCents = Math.round(totalCost * 100);
        totalCostCents.addAndGet(costCents);

        registry.counter("agent.tokens", "model", model, "type", "input")
            .increment(inputTokens);
        registry.counter("agent.tokens", "model", model, "type", "output")
            .increment(outputTokens);
        registry.counter("agent.cost.cents", "model", model)
            .increment(costCents);

        log.debug("Token usage - Model: {}, Input: {}, Output: {}, Cost: ${}",
            model, inputTokens, outputTokens, String.format("%.4f", totalCost));
    }

    public CostReport generateReport(Duration period) {
        // Query metrics for the period
        return CostReport.builder()
            .period(period)
            .totalCostCents(totalCostCents.get())
            .breakdown(getBreakdownByModel())
            .recommendations(generateCostRecommendations())
            .build();
    }

    private List<String> generateCostRecommendations() {
        List<String> recommendations = new ArrayList<>();

        // Analyze usage patterns and suggest optimizations
        // This is simplified - real implementation would analyze metrics

        recommendations.add("Consider using gpt-4o-mini for simple classification tasks");
        recommendations.add("Batch similar requests to reduce overhead");
        recommendations.add("Implement response caching for repeated queries");

        return recommendations;
    }

    private Map<String, Long> getBreakdownByModel() {
        // Implementation would query metrics
        return Map.of();
    }
}

record ModelPricing(double inputPricePerMillion, double outputPricePerMillion) {}

---

Orchestration Pattern Decision Matrix

Pattern	Latency	Quality	Cost	Complexity	Best For
Sequential	High	Medium	Low	Low	Simple pipelines, ETL
Hierarchical	Medium	High	Medium	High	Complex projects, research
Parallel	Low	High	High	Medium	Code review, multi-perspective
Consensus/Debate	Very High	Very High	High	High	Decision making, policy
ReAct	Medium	High	Medium	Medium	Tool-heavy tasks, QA
Plan-and-Execute	High	Very High	Medium	High	Complex multi-step tasks
Producer-Reviewer	High	High	Medium	Low	Content generation, code
Concierge	Low	Medium	Low	Low	Task routing, support

Best Practices Summary

1. Start Simple

Simple Task → Single Agent
Medium Task → Sequential Pipeline
Complex Task → Hierarchical + Specialists

2. Clear Agent Boundaries

• Each agent has ONE primary role
• Clear input/output contracts
• Limited context (need-to-know basis)
• Explicit handoff protocols

3. Observability First

• Trace every agent interaction
• Log structured events
• Monitor costs in real-time
• Alert on anomalies

4. Graceful Degradation

• Every operation has a timeout
• Fallback strategies for failures
• Circuit breakers for external dependencies
• Human escalation paths

5. Cost Awareness

• Choose right model for each task
• Cache repeated operations
• Batch where possible
• Monitor and optimize continuously

---

References

• Anthropic: "Building Effective Agents" (2024)
• Google DeepMind: "ReAct: Synergizing Reasoning and Acting" (2023)
• LangChain: "Multi-Agent Architectures" Documentation (2025)
• CrewAI: "Hierarchical Agent Systems" (2024)
• AutoGen: "Multi-Agent Conversation Patterns" (2024)
• Spring AI: "Agent and Tool Integration" Documentation (2025)
• Model Context Protocol (MCP): Official Specification (2024)
• OpenAI: "Best Practices for Agent Design" (2024)

---

Previous: 3.2 Multi-modal Prompting ← End of Series