Codex Harness is a workflow for running non-trivial agent tasks, in Codex or Claude Code, with evidence attached. It keeps that work from becoming a blur of confident summaries.

How a task runs

Each task is a small packet in a project’s harness folder:

  1. A state file names the task, its source of truth, the paths the agent may write to, the paths it must not touch, and the receipts it must produce.
  2. A preflight script validates the state file and checks the working tree before work starts.
  3. The agent works only inside the allowed paths and writes its status as a claim: blocked, in progress or complete.
  4. Each receipt is a JSON file pointing at evidence, such as a test log, build output, screenshot or source check.
  5. A verifier with clean context checks those files rather than the agent’s own account, and records a verdict: verified, verified with caveats, or not verified.
  6. Only verified output is promoted into project memory.

The point of the ceremony is to catch missing evidence before a task is reported as done. Later versions made the boundaries stricter: allowed write paths are compiled into an operating-system sandbox, protected-path changes are computed from the git diff, and hardcoded demo data in product code counts as a failure unless explicitly approved.

Use on this site

As of September 2026, work on this site still runs through the harness. The repository holds about 90 workstream folders, from the first room prototype in May 2026 through the move to Astro and the later project-showcase migrations. Most have their own state file, receipts and verifier report.