Automated Check★

An automated check is a deterministic test or validation run in the codebase environment (such as typecheckers, test suites, and linters). It returns a binary pass/fail result with precise console logs, allowing the agent to self-correct without human intervention.

You set up automated checks when you:

• Configure npm scripts like npm run test or npm run lint.
• Establish git hooks to block commits if typechecks fail.
• Setup CI/CD pipelines to run test suites on incoming branches.

The Self-Correction Loop

AI models are highly responsive to diagnostic error logs. If an agent writes code that contains a syntax error, it cannot detect it using reasoning alone. Instead, it relies on automated checks:

1. Write: The agent writes a code edit to a file.
2. Execute (Tool Call): The agent runs npm run build or tsc --noEmit as a tool call.
3. Log (Tool Result): The compiler returns a failure output: src/api/auth.ts:15:23 - Type 'string' is not assignable to 'number'.
4. Correct: The model reads the file name, line number, and error message, edits the target line, and runs the compile check again.

Agent Edits File ──► Runs Automated Check (e.g. Typecheck) ──► Error Output ──► Model Self-Corrects

For this loop to work, checks must be deterministic (identical code must always yield the identical result). A flaky test or unstable environment poisons the self-correction loop, causing the agent to break correct code or get stuck in loop cycles.

Field Applications & AX (Agent Experience)

The quality of automated checks is the single largest factor in a repository's AX (Agent Experience). A project with strict TypeScript typing, instant lint checks, and fast unit tests enables the agent to catch and fix 95% of its own errors before you ever see them.