Multi Agent Plan Review

multi-agent-plan-review

Cross-agent pre-implementation plan review and repair. Dispatch review agents for independent critique; keep judgment, synthesis, artifact revision, and readiness decisions with the current orchestrator.

Required inputs

Fail before side effects unless all are present:

• review_agents: one or more dispatch agent roles
• plan artifact path
• task artifact path
• context artifact path

Optional input:

• spec artifact path, used as the review contract when present

Dispatch agent roles may be terse (codex high, claude opus xhigh, antigravity). Infer the agent only when unambiguous. Before dispatching, follow agent-dispatch, inspect live options, and fail if a requested override is unsupported.

Required artifacts

Use run-id = YYYYMMDD-HHMMSS-:

• .agent-layer/tmp/multi-agent-plan-review..state.md
• .agent-layer/tmp/multi-agent-plan-review..report.md

Create both before writing. Store review agent prompt/output artifacts under .agent-layer/tmp/ with the same prefix and record each path in state.

Rules

• Review against the spec when a spec path is supplied; otherwise review against

the plan's stated objective and scope.

• Use review-plan only for dispatched review agent runs. Do not ask review agents to

edit artifacts.

• Treat review agent output as suspect. Accept a suggestion only when you agree with

it.

• Classify each finding as accepted, rejected, duplicate, or

substantive-user-decision; record a one-line reason.

• This is not a report-only review: revise the plan/task/context artifacts for

accepted findings and resolved user decisions.

• Revise artifacts only for accepted findings and resolved user decisions.
• Ask the user before applying a substantive change or tradeoff not settled by

the spec or plan. Use concrete options with brief pros, cons, and a recommendation.

• Iterate until no accepted unresolved findings remain and no substantive user

decision is pending.

• Dispatch review agents one at a time unless the dispatch implementation is known

to be safe for concurrent review agent launches in the current repo.

Workflow

Phase 1: Preflight

1. Read the plan, task, and context artifacts fully.
2. Read the spec if one was supplied.
3. Confirm the artifacts describe the same objective, scope, and execution

contract.

1. Normalize dispatch agent roles through live al dispatch options.
2. Record artifact paths, dispatch agent roles, normalized dispatch flags, and

current round in the state file.

Phase 2: Dispatch review agents

For each review agent dispatch role, dispatch review-plan with a focused prompt containing:

• plan, task, context, and optional spec paths
• instruction to review only the artifact set
• instruction to report blockers, weak verification, missing scope, sequencing

gaps, hidden assumptions, and docs/tests/memory gaps

• instruction to avoid rewriting the plan

Capture the useful review agent result or report path in the state file.

Phase 3: Synthesize

For every review agent finding:

1. Validate it against the artifacts and repo context.
2. Classify it under Rules.
3. Ignore speculative, unsupported, out-of-scope, or merely stylistic findings.
4. Merge duplicates before revising artifacts.

Phase 4: Revise accepted findings

Keep revisions scoped to implementation readiness:

• clarify scope or non-goals
• fix sequencing
• strengthen verification
• add missing docs/tests/memory steps
• update context paths, entry point, or constraints

Do not convert rejected suggestions into churn.

Phase 5: Repeat to alignment

After any artifact revision, repeat Phases 2-4.

Phase 6: Report

Write the final report with these sections:

1. # Multi-Agent Plan Review Summary
2. ## Inputs
3. ## Review Agent Rounds
4. ## Accepted Changes
5. ## Rejected Suggestions
6. ## User Decisions
7. ## Final Readiness

Final Readiness must be exactly one of:

• implementation-ready
• blocked-for-user-decision

Guardrails

• Do not treat review agent consensus as truth or hide rejected suggestions.
• Do not widen implementation scope or weaken verification to satisfy a

review agent preference.

Definition of done

• All required artifacts and dispatch agent roles were validated.
• Every review agent dispatch role ran through review-plan.
• Every review agent finding was classified with a reason.
• Accepted findings were resolved in the artifacts or escalated through a human

checkpoint.

• The final report exists and declares implementation-ready or

blocked-for-user-decision.

Final handoff

Echo the report path, the reviewed plan/task/context paths, the review agent dispatch roles used, accepted changes made, rejected suggestions, and final readiness.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: conn-castle
• Source: conn-castle/agent-layer
• License: MIT
• Homepage: https://agent-layer.dev/

Install and usage instructions live in the source repository linked above.