Multiple agents: divide work only when the boundary helps

Three agents agree on an answer. Have you gained three independent pieces of evidence? Only if their work adds independent support. Start by mapping the data and deliverable each worker contributes.

Before you begin: You understand a tool-using agent and shared application state.

An agent here is a model-driven component with a role, context, and permitted actions. Several agents can use the same underlying model. Giving them different names does not make their errors independent.

Find a task that can actually be divided

Suppose a research request needs information from three independent document collections. Separate workers can search those collections in parallel and return evidence. A coordinator can then combine results and identify gaps.

By contrast, asking a planner, thinker, critic, and writer to repeatedly discuss a simple factual question may add little. A single model with a read-only retrieval tool and an explicit output contract is a useful baseline.

Choose agent boundaries around different data access, tools, expertise requirements, or independently checkable deliverables. Boundaries based only on personalities are difficult to evaluate.

Define a handoff as data

A worker should return the task identifier, supported findings, source references, unresolved questions, and status. A coordinator should not have to infer completion from a paragraph that sounds confident.

Pass only the context needed for the task. Copying the entire conversation and all credentials to every worker increases cost and access risk. Each tool still enforces the acting user's permissions.

Give shared state one clear owner

Parallel workers can overwrite each other's updates if they all edit a single mutable plan. Prefer separate results keyed by task ID and a coordinator that resolves conflicts deliberately. Use durable state when work must survive restarts.

A worker timeout does not mean its external action never happened. If a worker can change state, use operation identities and recovery procedures. Avoid giving both coordinator and worker responsibility for the same side effect without a clear protocol.

Two agents agreeing is not equivalent to two independent sources agreeing. They may share the same model, prompt, retrieved document, or misconception. A critic adds value when it checks evidence or a concrete invariant, not merely when it produces another opinion.

Remove one worker and predict what disappears

Suppose a coordinator has a catalog reader, a date checker, and a writer. The reader returns records; the checker validates date consistency against those records; the writer produces a response. Remove the checker and ask which failure should become more likely. If you cannot name one, its role may be too vague to evaluate.

Keep worker outputs separately so the coordinator can identify the source of a conflict. Then measure both total model usage and elapsed time. Parallel work can improve one while worsening the other; agreement alone measures neither.

Measure the system against a simpler baseline

Compare task success, unsupported claims, latency, total cost, and handoff failures. Parallel work can reduce elapsed time while increasing total model usage. Include unsuccessful tasks in cost accounting rather than reporting only cheap successful examples.

Exercise: three agents independently repeat a false fact found on the same webpage. Does majority voting increase confidence?

Next, build a research workflow whose outputs can be checked claim by claim.

Sources

The LangChain multi-agent documentation describes orchestration patterns. The A2A specification addresses communication between agents; neither source implies that extra agents always improve a task.

Continue: Project: Autonomous Research Agent.