multi-agent-orchestration

Multi-Agent Orchestration

Use multiple agents only when specialization or safe parallelism outweighs coordination cost.

Use when

• Split a large objective into independent, verifiable workstreams.
• Coordinate specialists that need distinct tools, permissions, or context.
• Run worker-reviewer, planner-executor, map-reduce, or bounded debate patterns.
• Diagnose duplicate work, conflicting edits, weak handoffs, or stalled dependencies.

Do not delegate a tightly coupled, small, or inherently sequential task merely to increase agent count.

Inputs

Collect the objective, completion criteria, task graph, available agents and tools, concurrency limits, shared files or systems, authority boundaries, deadlines, budget, and final decision owner. State assumptions and unresolved dependencies.

Output contract

Produce:

1. A decomposition rationale and explicit non-goals.
2. A directed acyclic task graph with owner, dependencies, inputs, output contract, write scope, and verification for every task.
3. A handoff protocol and shared-state policy.
4. Approval points, timeout and retry limits, escalation routes, and stop conditions.
5. A synthesis plan that resolves disagreements and verifies the integrated result.
6. A completion report with evidence, remaining uncertainty, and unused or failed branches.

Workflow

1. Define one measurable objective and the authority boundary before assigning work.
2. Decompose by separable outputs, not vague roles. Keep shared mutable state to a minimum and retain tightly coupled steps under one owner.
3. Draw dependencies and identify the critical path. Parallelize only tasks with independent inputs and non-overlapping side effects. Read orchestration-patterns.md when choosing a topology.
4. Assign one accountable owner per task. Specify inputs, deliverable format, write scope, validation, deadline or timeout, and what warrants escalation.
5. Give each agent the minimum context and permissions needed. Include source artifacts, not hidden conclusions, when independent judgment matters.
6. Require structured handoffs: status, result, evidence, changed state, assumptions, risks, and next dependency. Acknowledge receipt before downstream mutation.
7. Monitor dependency state and useful progress. Bound retries and debates; do not recursively delegate without a clear capacity and ownership model.
8. Synthesize centrally or through a named integrator. Resolve conflicting claims from primary evidence, run integration checks, and confirm the original completion criteria.
9. Close or cancel unused work, record unresolved risks, and return control to the final decision owner.