Agent Output Review Gate (QA)

Agent Output Review Gate (QA)

1. 01

   Freeze the agent session log before anyone edits the branch

   Export the full session transcript and diff so the review is judged against what the agent actually did, not what a developer remembers it doing. Tools such as Agent Monitor and the open-source Agent Desk exist specifically to browse sessions across Claude Code, Cursor, Codex, and Copilot.

2. 02

   Score the diff against the client's architectural constraints, not just test pass rate

   A passing suite says nothing about whether the agent introduced a second state library or bypassed the auth layer. The Parsimony benchmark was built on the premise that solve rate alone hides code bloat, and the same gap shows up in agency delivery.

3. 03

   Run the silent-failure sweep on anything the agent generated for a client-facing surface

   Bugs that never throw are the ones that reach production: text matching its background color, client-side race conditions on form submit. A 10 MB in-browser model such as GenClass can flag these without a server round-trip.

4. 04

   Confirm every external API contract the agent touched still matches its consumer code

   When a spec changes, the breakage lands in repositories the agent never opened. Ripple detects the change and opens pull requests against affected consumers across OpenAPI, Protobuf, GraphQL, and database schemas.

5. 05

   Assign a named human owner to the architectural decisions inside the diff

   The category description is explicit that these tools scaffold and maintain; they do not replace architectural oversight. If no one on the retainer can defend a design choice in the diff, it does not ship.

6. 06

   Log the review outcome and the agent's miss pattern in the project's shared reference

   Track which failure classes recur (auth bypass, duplicated utilities, unhandled nulls) so the next prompt template closes the gap instead of repeating it.