Reviewer & QA agents

An agent that just wrote something is soaked in its own reasoning. A reviewer with a clean context reads the output cold and often catches things the writer skipped. But it's the same kind of model: it can share the same blind spots and the same wrong facts. So give it what makes it useful — the original ask, the source files, and explicit checks — and test it: plant a couple of known mistakes and see if it finds them. For anything with money, legal, or a client on it, a person still checks. Two more traps. Models lean toward agreeing with you, because they were trained on what people liked to hear. Ask 'is this good?' or say you wrote it, and the review comes back kinder. So ask neutrally: 'review this against these checks,' and don't say whose work it is. And a review is a judgement, not a measurement: run the same reviewer twice and it can flag different things. Treat one clean review as one opinion.

The long version: Reviewer subagents ↗

Checkpoints

• ▸You should now see a reviewer spec that reads the deliverable, the original ask, and the sources, and reports issues only.
• ▸You should now see whether the reviewer found your planted mistakes.
• ▸You should now see a list of findings you sorted into real and false alarms.