Multi-agent systems, handoffs, and human approval

Multiple agents can specialize, parallelize independent research, or provide separation of duties. They also multiply latency, cost, context loss, security surface, and nondeterminism.

Useful patterns

• supervisor routes to bounded specialist agents;
• fan-out retrieves independent sources, then deterministic merge;
• reviewer evaluates an artifact against a rubric;
• producer/critic loop with a strict iteration cap;
• handoff transfers explicit typed state and evidence.

Do not create one agent per department merely to draw an impressive diagram.

Handoff contract

Use a schema containing task, known facts, source IDs, actions completed, uncertainties, remaining budget, and next authority. Do not rely on a prose summary as the only state.

Human-in-the-loop

Pause before payments, messages, record mutations, privilege changes, policy exceptions, medical/legal conclusions, or irreversible work. The approval UI must show exact proposed action, parameters, evidence, risk, and alternatives. Approval must be scoped and expire.

Separation of duties

The agent proposing an action should not be the sole validator for high-risk work. Use deterministic rules, independent policy service, or authorized human.

Failure containment

Give each agent the minimum tools and data. Use per-agent budgets, deadlines, circuit breakers, and trace IDs. Cancel siblings when their output is no longer needed.

Feynman check

Adding agents is hiring more junior operators. They need job boundaries, handoff forms, supervisors, budgets, and a rule for when a human must sign.