What is Revue?
Revue is the decision plane for teams running coding agents. Its resident manager lets you supervise a swarm from one Amp thread: repeated blockers become one human decision, each answer returns to the waiting sessions, and completed children wake the manager so work can continue.
The thesis is simple: human judgment is the scarcest resource in an agent swarm. Agents should resolve routine implementation details themselves. When several agents genuinely need the same judgment, the human should answer once.
One decision instead of three interruptions
Imagine three agents working in parallel:
- Agent A asks whether to keep a callback compatibility layer.
- Agent B reaches the same fork while changing a neighboring handler.
- Revue groups both asks into one decision and shows who is waiting, the options, the recommendation, and the risk.
- You answer once. Revue routes that answer to both agents.
- Agent C hits the same fork tomorrow. Revue applies the remembered answer and records that another interruption was prevented.
Revue did not make the product decision. It compressed one human judgment, delivered it to every dependent session, and preserved proof of what happened.
How Revue protects attention
| Stage | Revue behavior |
|---|---|
| Search memory | An agent checks prior human decisions before asking again |
| Compress asks | Near-duplicate blockers from different agents join one decision |
| Improve the interruption | The decision carries options, recommendation, risk, evidence, and affected scope |
| Route one answer | Every contributing workstream receives the same human decision |
| Prove and reuse | Receipts show what was unblocked or suppressed; corrections keep memory honest |
Read How a decision moves for the full lifecycle and the limits of each receipt.
Product shape
Revue has one decision-plane implementation with local and hosted deployments. The local plane is account-free and loopback-only. The hosted plane uses browser approval and supports agents that do not share your machine. Their decision memory is separate.
The resident Amp manager is the preferred cockpit, not the product boundary. revue serve keeps it running through amp --no-tui; cleared blockers and child completions arrive in the manager thread without requiring you to watch every child. Claude Code, Codex, Cursor, OpenCode, and other MCP harnesses ask through the same decision plane with revue-mcp. Humans can answer from the manager or the terminal inbox.
When Revue fits
Use Revue when:
- several agents work concurrently and may reach overlapping product or architecture forks;
- repeated human decisions are expensive or easy to apply inconsistently;
- an answer must return to a running agent rather than remain in a separate dashboard;
- you need an audit trail for where a decision landed and when it was reused.
Revue is probably unnecessary when one short-lived agent can complete the work without human judgment, or when a repository already answers the question unambiguously. Revue is not a substitute for reading the codebase, running tests, or exercising engineering judgment.
Start here
- Follow the Quickstart to see the complete loop and connect one agent.
- Read Local vs hosted before choosing where decision memory lives.
- Use Connect an MCP harness for Claude Code, Codex, Cursor, OpenCode, or another MCP client.
- Use Run the Amp manager to supervise Amp children, Orbs, and subagents.