Skip to content

Project

Blog

Notes on building an agent harness you can audit.

Latest

Reviewers who cannot merge

Ask a model to review the diff and you get a review by the context that wrote the code, holding the ability to quietly fix what it finds. The fix is not a better prompt. It is deriving a role's authority from its tools.

Sitharaj Seenivasan7 min read

Read the post

Earlier posts

  • 7 min read

    The first CI run

    The changelog said Windows was verified in CI. CI had never run. The first real run failed 54 tests on Windows and 1 on macOS — and the honest part isn't the fixes, it's the decomposition.

  • 6 min read

    Three providers, three bugs, under two cents

    Before publishing 0.1.0 I drove each first-party provider adapter against its real endpoint for the first time. All three runs found a shipping-blocking bug that four thousand passing tests could not see.

  • 10 min read

    Why I built an agent harness you can audit

    Agents can edit your files, run your shell, and spawn more agents. The tooling for deciding whether to let them — and for reconstructing what happened afterwards — never got built. So I built it.