orchestratelisted

# AEP — Orchestration Protocol **Hard rule: a subagent's report is a claim, not a result.** In multi-agent work the dominant failure mode is inter-agent misalignment — agents confidently relaying wrong findings and validating each other's errors into "shared truth". The lead verifies; it never just collects. ## 1. Decompose like a lead engineer - Split by **seam, not by size**: independent modules, independent questions, independent verification lenses. Two subagents editing the same file is a decomposition failure. - Each unit of work must be *independently completable and independently checkable* — if you cannot state how you will verify a unit without seeing the others, merge or re-cut the units. - Keep for yourself what cannot be delegated: goal ownership, integration, final verification, and anything requiring cross-unit judgment. ## 2. Brief with a contract Every delegation carries a written contract: ``` Goal: <one line — what done means> Inputs: <files/paths/context the subagent needs — no more> Acceptance: <numbered, checkable criteria — the integration test for this unit> Out of scope:<what it must NOT touch> Return: <exact format expected back: findings + evidence, diff, list> ``` Vague briefs produce confident garbage. If you cannot write the acceptance criteria, you do not understand the unit yet — return to `aep:explore`. ## 3. Isolate contexts deliberately - Give each subagent what its contract needs and nothing else. Context isolatio