multi-agent-patterns

name	multi-agent-patterns
description	Multi-Agent Systems: orchestration vs choreography, tool routing, state management, agent handoffs, parallelization (fan-out/fan-in), error handling in multi-agent workflows, Claude SDK patterns (Agent/Tool/Handoff), and observability with OpenTelemetry.

Multi-Agent Patterns

Patterns for building reliable, scalable multi-agent systems with Claude.

When to Activate

• Designing a system where multiple Claude agents collaborate
• Implementing task decomposition with parallel sub-agents
• Routing tasks between specialized agents (router/dispatcher patterns)
• Managing state across agent invocations
• Building reliable agent-to-agent handoffs
• Adding observability (tracing, latency) across agent boundaries
• Selecting the right coordination pattern (fan-out, pipeline, split-role)

---

Orchestration vs. Choreography

Orchestration (Central Control)          Choreography (Decentralized Events)
─────────────────────────────────        ─────────────────────────────────────
        Orchestrator                      Agent A ──event──▶ Agent B
       /     |     \                      Agent B ──event──▶ Agent C
   Agent A  Agent B  Agent C             (no central coordinator)

WHEN: Clear workflow, sequential steps   WHEN: Loose coupling, event-driven, scaling
      easy to debug and reason about            independent microservices

Choose Orchestration when:

• Workflow steps are known in advance
• You need a single audit trail
• Failure handling requires central coordination
• Order of execution matters

Choose Choreography when:

• Services must remain independent
• New consumers can subscribe without modifying producers
• Eventual consistency is acceptable
• Scale requires horizontal distribution

---

Tool Routing

The orchestrator decides which agent/tool handles a task.

Intent Classification Router

// Claude as router — classify then dispatch
const AGENT_REGISTRY = {
  'code-review': codeReviewAgent,
  'security-scan': securityAgent,
  'test-generation': tddAgent,
  'documentation': docAgent,
};

async function route(task: string, context: string): Promise<AgentResult> {
  const classification = await claude.messages.create({
    model: 'claude-haiku-latest',  // Haiku for lightweight routing
    system: `Classify the task into one of: ${Object.keys(AGENT_REGISTRY).join(', ')}.
             Reply with ONLY the category name.`,
    messages: [{ role: 'user', content: task }],
    max_tokens: 10,
  });

  const category = classification.content[0].text.trim();
  const agent = AGENT_REGISTRY[category];

  if (!agent) throw new Error(`No agent for category: ${category}`);
  return agent.(task, context);
}

---

State Management

Where does state live between agent calls?

In-Memory (Short-Lived Workflows)

interface WorkflowState {
  taskId: string;
  input: string;
  steps: StepResult[];
  metadata: Record<string, unknown>;
}

class WorkflowContext {
  private state: WorkflowState;

  constructor(taskId: string, input: string) {
    this.state = { taskId, input, steps: [], metadata: {} };
  }

  addStep(name: string, result: unknown): void {
    this.state.steps.push({ name, result, timestamp: Date.now() });
  }

  getLastResult(): unknown {
    return this.state.steps.at(-1)?.result;
  }

  toHandoffSummary(): string {
    // Compress context for sub-agent handoffs
      +
       +
      ;
  }
}

For durable state (Redis, DynamoDB event log) and task decomposition handoffs, see multi-agent-patterns-advanced.

---

Handoff Patterns

Full Context Handoff

Pass the complete conversation history to the next agent — use when the sub-agent needs full context.

async function handoffWithFullContext(
  conversation: Message[],
  nextAgentSystem: string
): Promise<string> {
  const response = await claude.messages.create({
    model: 'claude-sonnet-latest',
    system: nextAgentSystem,
    messages: conversation,  // Full history passed through
    max_tokens: 4096,
  });
  return response.content[0].text;
}

Summary Handoff

Compress context before handoff — use for long workflows or to save tokens.

async function summarizeForHandoff(
  context: WorkflowContext,
  maxTokens = 500
): Promise<string> {
  const summary = await claude.messages.create({
    model: 'claude-haiku-latest',  // Haiku for cheap summarization
    system: 'Summarize the key findings and decisions. Be concise.',
    messages: [{
      role: 'user',
      content: `Summarize this workflow progress for handoff to next agent:\n${context.toHandoffSummary()}`,
    }],
    max_tokens: maxTokens,
  });
  return summary.content[0].text;
}

---

Parallelization (Fan-Out / Fan-In)

async function parallelReview(codeFiles: string[]): Promise<ReviewResult[]> {
  // Fan-out: launch all reviews concurrently
  const reviewPromises = codeFiles.map(file =>
    reviewAgent.run(file).catch(err => ({
      file,
      error: err.message,
      issues: [],
    }))
  );

  // Fan-in: collect results
  const results = await Promise.allSettled(reviewPromises);

  return results.map((result, i) => {
    if (result.status === 'fulfilled') return result.value;
    return { file: codeFiles[i], error: result.reason.message, issues: [] };
  });
}

// Controlled parallelism (avoid rate limits)
async function parallelWithConcurrencyLimit<T>(
  tasks: (() => Promise<T>)[],
  concurrency = 5
): Promise<T[]> {
  const results: T[] = [];
   chunks = [];

   ( i = ; i < tasks.; i += concurrency) {
    chunks.(tasks.(i, i + concurrency));
  }

   ( chunk  chunks) {
     chunkResults =  .(chunk.( ()));
    results.(...chunkResults);
  }

   results;
}

---

Error Handling in Multi-Agent Systems

Retry with Fallback Agent

async function runWithFallback<T>(
  primary: () => Promise<T>,
  fallback: () => Promise<T>,
  maxRetries = 2
): Promise<T> {
  for (let attempt = 0; attempt <= maxRetries; attempt++) {
    try {
      return await primary();
    } catch (err) {
      if (attempt === maxRetries) {
        console.warn(`Primary agent failed after ${maxRetries} retries, using fallback`);
        return fallback();
      }
      await backoff(attempt);
    }
  }
  throw new Error('Unreachable');
}

// Partial results: continue with what succeeded
async function collectPartialResults<T>(
  tasks: Promise<T>[],
  minRequired: number
): Promise<T[]> {
  const results = await Promise.allSettled(tasks);
  const successes = results
    .filter((r): r is PromiseFulfilledResult<T> => r.status === 'fulfilled')
    .( r.);

   (successes. < minRequired) {
      ();
  }
   successes;
}

---

Claude Agent SDK Patterns

The core agentic loop: call the model, handle tool_use stop reason by executing tools and appending results, repeat until end_turn.

import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic();

async function orchestratorLoop(goal: string): Promise<string> {
  const messages: Anthropic.MessageParam[] = [{ role: 'user', content: goal }];

  while (true) {
    const response = await client.messages.create({
      model: 'claude-sonnet-latest',
      system: ORCHESTRATOR_SYSTEM_PROMPT,
      tools: AVAILABLE_TOOLS,
      messages,
      max_tokens: 4096,
    });

    if (response.stop_reason === 'end_turn') {
      return response.content.filter(b => b.type === 'text').map(b => b.text).join('');
    }

    // Handle tool calls: execute all in parallel, append results
    messages.({ : , : response. });
     toolResults =  .(
      response.
        .((b): b is . => b. === )
        .( (t) => ({
          :   ,
          : t.,
          :  (t., t.),
        }))
    );
    messages.({ : , : toolResults });
  }
}

---

Observability

import { trace, context, SpanStatusCode } from '@opentelemetry/api';

const tracer = trace.getTracer('multi-agent-system');

async function tracedAgentCall<T>(
  agentName: string,
  task: string,
  fn: () => Promise<T>
): Promise<T> {
  return tracer.startActiveSpan(`agent.${agentName}`, async (span) => {
    span.setAttributes({
      'agent.name': agentName,
      'agent.task.length': task.length,
      'agent.task.preview': task.slice(0, 100),
    });

    try {
      const result = await fn();
      span.setStatus({ code: SpanStatusCode.OK });
      return result;
    } catch (err) {
      span.setStatus({ code: SpanStatusCode.ERROR, message: String(err) });
      span.recordException(err as Error);
       err;
    }  {
      span.();
    }
  });
}


 ():  {
  .(.({
    : ,
    ...event,
    :  ().(),
  }));
}

---

Anti-Patterns

Running Independent Sub-Agents Sequentially Instead of in Parallel

Wrong:

// Runs one at a time — wastes latency when tasks are independent
const reviewResult = await codeReviewAgent.run(code)
const securityResult = await securityAgent.run(code)
const docsResult = await docAgent.run(code)

Correct:

// Fan-out: all three run concurrently
const [reviewResult, securityResult, docsResult] = await Promise.all([
  codeReviewAgent.run(code),
  securityAgent.run(code),
  docAgent.run(code),
])

Why: Sequential execution of independent agents multiplies latency unnecessarily — fan-out with Promise.all reduces wall-clock time to the slowest single agent.

---

Passing the Full Conversation History to Every Sub-Agent

Wrong:

// Sub-agent receives 50 turns of unrelated orchestrator history
async function handoff(fullConversation: Message[], nextSystem: string) {
  return claude.messages.create({
    system: nextSystem,
    messages: fullConversation,  // bloats context, raises cost, degrades focus
    max_tokens: 4096,
  })
}

Correct:

// Compress to only what the next agent needs
const summary = await summarizeForHandoff(context, 500)
return claude.messages.create({
  system: nextSystem,
  messages: [{ role: 'user', content: summary }],
  max_tokens: 4096,
})

Why: Passing irrelevant history inflates token costs, risks hitting context limits, and distracts sub-agents with information they don't need.

---

Using Opus for Lightweight Routing Decisions

Wrong:

const classification = await claude.messages.create({
  model: 'claude-opus-latest',   // overkill for a one-word classification
  system: 'Classify as: code-review | security-scan | documentation.',
  messages: [{ role: 'user', content: task }],
  max_tokens: 10,
})

Correct:

const classification = await claude.messages.create({
  model: 'claude-haiku-latest',  // fast and cheap for routing
  system: 'Classify as: code-review | security-scan | documentation. Reply with the category only.',
  messages: [{ role: 'user', content: task }],
  max_tokens: 10,
})

Why: Routing is a lightweight classification task; using a heavyweight model wastes cost and latency on a decision that requires no deep reasoning.

For advanced patterns — capability registry, durable state, task decomposition, testing multi-agent systems, pattern quick-selection guide, and failure handling — see multi-agent-patterns-advanced.