← All field notes
02 / Orchestration

When a multi-agent system earns its complexity.

Specialization can expand what an AI system can do. It also creates a new job: coordinating the workers, their context, and the evidence they return.

Direct answer

A multi-agent system is not automatically an AI team. Without clear ownership, bounded context, and a synthesis mechanism, it is one model creating more places for the work to go wrong.

Orchestrator-worker patternOne objective · specialist ownership
Lead orchestratorplans · delegates · synthesizes
Plannerscope + criteria
Architectsystem boundaries
Builderreviewable change
Verifierindependent evidence
SPEC ✓
DESIGN ✓
DIFF ✓
TESTS ✓
The orchestrator should synthesize verifiable artifacts—not trust confident summaries from its workers.

The architecture becomes useful when a task contains genuinely different lines of work that can be delegated. Anthropic’s production research system uses an orchestrator-worker pattern: a lead agent develops a strategy, creates specialized subagents, and synthesizes what they return. The pattern fits research because the work is broad, parallelizable, and difficult to plan completely in advance.

The three signs of a good fit

Multi-agent orchestration earns its cost when at least one of these conditions is true:

  1. The task decomposes into independent searches or analyses. Parallel workers can cover more ground without blocking one another.
  2. Different roles need different context or tools. A security reviewer should not carry the same instructions and permissions as a deployment agent.
  3. The plan must adapt as evidence arrives. An orchestrator can create new work when a worker returns a gap, contradiction, or unexpected dependency.

If every worker needs the full context, changes the same artifact, and waits for the previous worker, a multi-agent label may only be hiding a sequential workflow.

The handoff is the product

The quality of a multi-agent system is often determined between agents. Every delegation needs a contract: the objective, available context, permitted tools, expected output, success criteria, and what to do when blocked.

Anatomy of a handoffEvery delegation is a contract
Objectivethe owned outcome
Contextonly what is needed
Toolsbounded authority
Criteriadefinition of done
Escalationwhat to do when blocked
Evidenceartifact the next role can verify
Weak handoffs transfer prose. Strong handoffs transfer responsibility, constraints, and proof.

Workers should return evidence, not confidence. A coding worker can return a diff and test results. A research worker can return claims mapped to sources. A deployment worker can return resource identifiers and health checks. These artifacts let the orchestrator synthesize from verifiable work instead of trusting a persuasive summary.

Design ruleGive each worker the minimum context required to own its result—and require enough evidence for the next agent or human to verify it.

The hidden costs of “more agents”

Coordination creates failure modes a single loop does not have:

  • Duplicated work: vague assignments cause workers to explore the same path.
  • Context loss: a summary drops the constraint that mattered downstream.
  • Conflicting edits: workers modify shared state without ownership rules.
  • Error amplification: the lead agent treats a weak worker result as established fact.
  • Unbounded spend: agents spawn more work without a budget or stopping condition.
Coordination failure chainOne vague task can compound downstream
Vague assignmentownership overlaps
Duplicate workcontext diverges
Weak synthesiserrors become facts
Wrong outcomehigh cost · low traceability
Coordination failures are architectural. A more elaborate persona prompt does not repair them.

These are systems problems. Solve them with task identifiers, explicit state, concurrency limits, tool permissions, durable artifacts, trace logs, and escalation rules—not with a more elaborate role prompt.

A production-shaped organization

For product engineering, a useful organization maps roles to distinct responsibilities rather than fictional personalities:

PlannerOwns scope, assumptions, risks, and acceptance criteria.
ArchitectOwns boundaries, data, interfaces, and tradeoffs.
BuilderOwns implementation and a reviewable change set.
VerifierOwns tests, evidence, and independent challenge.

The deploy role should not merely run a command. It should verify environment, version, health, rollback readiness, and the approval state. When responsibilities are explicit, orchestration can enforce separation of concerns instead of simulating a busy group chat.

Evaluate the organization, not just the answer

End-to-end quality matters, but it does not explain why a run succeeded or failed. Capture the full trajectory: the plan, delegated tasks, tool calls, artifacts, corrections, cost, time, and final outcome.

Then test coordination itself. Did the orchestrator divide the task cleanly? Did it provide enough context? Did it recognize a bad result? Did workers stay inside their authority? Could a human reconstruct the important decisions?

A multi-agent system is justified when it produces better outcomes on representative work after coordination costs are included. Until that is measured, “more agents” is an architecture hypothesis.

Sources and further reading

  1. Anthropic — How we built our multi-agent research system
  2. Anthropic — Building effective agents
  3. Anthropic — Effective context engineering for AI agents
Next field noteHuman gates are part of the architecture.
Continue to note 03 →