Automated check

An automated check is anything that can say yes or no about the agent's work without a person looking: the type checker, the linter, the test suite, a build, a schema validation, a link checker. It has a definite answer and it runs in seconds or minutes. It is the cheapest feedback an agent can get, and unlike a human it never gets tired of running.

Checks matter more with agents than with people because the agent will happily continue past a mistake it can't see. A hallucinated method name compiles in the agent's head but not in the compiler. Without a check, that error surfaces three files later as a confusing failure; with one, the agent gets a tool result saying exactly what broke and fixes it in the next turn.

Wire the checks into the loop rather than running them yourself at the end. Tell the agent in AGENTS.md which command to run after every change, or use hooks to run it automatically after each edit. The stricter and faster your checks, the further you can let the agent run AFK. What checks can't judge (taste, product sense, whether the feature is the right feature) is the job of human review.