Multi-agent orchestration

Managing bounded contexts across complex subnetworks and localized state interfaces.

Why it matters

Monolithic agents breakdown under massive, complex prompts. Multi-agent orchestration resolves this by isolating specific tasks into dedicated, isolated sub-graphs. By sharing global state properties while keeping internal node histories localized, you maximize token efficiency and significantly stabilize model accuracy.

Multi-agent orchestration splits a hard task across specialised sub-agents (researcher, analyst, critic) that share global state but keep their internal histories local. The goal is bounded context per role — each agent sees only what it needs, which improves accuracy and token efficiency.

supervisor:
  state: [question, plan, evidence, answer, risk]
  agents:
    researcher: retrieves chunks and graph paths
    analyst:    turns evidence into claims
    critic:     flags unsupported claims and policy risk
  handoff_rule: only the supervisor mutates the final answer
  stop_condition: critic finds no unsupported claims OR max_rounds == 3

The critic role is what separates multi-agent from parallel prompting. Without an agent whose job is to verify, you've just run several models at once and concatenated the results. The critic checks claims against evidence and sends the answer back for revision — that verification loop is the whole point.

Shared state beats message-passing because every node sees the full, current context without re-serialising and re-summarising it between hops (which loses information). One state object, mutated by the supervisor, keeps a single source of truth.

Bound the loop and the budget. max_rounds stops infinite researcher↔critic ping-pong, and a budget field decides which agent gets cut first under cost pressure. Coordination without stop conditions is how multi-agent systems burn money.

Multi-agent orchestration is a newsroom: reporters gather facts, an editor shapes them, and a fact-checker can kill the story until claims are sourced. Drop the fact-checker and you don't have a faster newsroom — you have rumours published in parallel.

• No critic/verifier. Symptom: unsupported claims ship. Fix: a dedicated verification agent + revise loop.
• Message-passing everything. Symptom: context loss between hops. Fix: shared state object.
• No stop condition. Symptom: infinite rounds. Fix: max_rounds / convergence check.
• No budget rule. Symptom: cost blowups. Fix: decide which agent is cut first under pressure.