references/multiagent-orchestration.md

# Multi-agent orchestration contracts
Start with cost comparison against a single coordinator and deterministic subprocesses. Delegate only independent specialized tasks. Manager owns final accountability; each worker has one scoped job, permitted data/tools, output schema, and cost cap. Read-only reviewers should not edit implementation files. Implementation workers use nonoverlapping paths/branches/worktrees.

Parallelize reference checking, independent reviews and isolated implementation components. Do not parallelize competing edits or external writes. Bound concurrency, retries and circular handoffs. A specialist report includes: scope, artifacts, commands actually executed, evidence, unresolved questions, security findings. Manager reconciles evidence, never counts verbal affirmation as proof.

**Truth rule:** Skills do not create real subagents by themselves. If the runtime lacks subagent execution, perform the planned roles sequentially and label them accordingly. For current SDK semantics see https://openai.github.io/openai-agents-python/multi_agent/ and https://openai.github.io/openai-agents-python/handoffs/.