Skip to main content

Recipe evidence loop

The recipe loop turns agentic work from “the diff looks plausible” into “the behavior was exercised and the proof is reviewable.”

Loop​

Why this matters​

Without recipes, agents repeatedly explore the same UI and humans still need to guess whether validation was meaningful.

With recipes, validation becomes a reusable asset:

  • a bug reproduction recipe can prove the fix;
  • an acceptance-criteria recipe can accompany a PR;
  • a regression recipe can run later against adjacent changes;
  • evidence can be attached to review instead of described in prose.

Evidence package​

A compatible runner should produce a typed artifact package containing at least:

  • summary.json — high-level run result;
  • trace.json — ordered action trace;
  • artifact-manifest.json — typed index of screenshots, videos, logs, and other outputs;
  • resolved recipe/workflow copy when practical.