Harness²: The Agent Harness That Rewrites Itself

A harness is everything you wrap around a coding agent so it behaves like a teammate: standing instructions, permissions, hooks, tools and a verification loop. Most people build one, feel the improvement, and then freeze it. Six weeks later the file is stale, the agent is repeating last month’s mistakes and nobody remembers which rule was written for what.

The short from DIY Smart Code titled “Why your AI agent needs to rewrite its own rules” (tagged #harness2) names the next step. This is our longer version of that idea: the harness needs its own harness. Call it harness².

Why a static harness decays

Codebases move. Dependencies change, conventions drift, a deploy quirk gets fixed. Every rule in a memory file costs context on every session, and a contradicted rule is worse than none — it teaches the agent that the file is optional. A harness that only grows becomes noise. A harness that never grows keeps paying for the same mistake.

The model above shows the shape, not a measurement: promotion collapses repeat mistakes quickly, while pruning keeps the rule set from swelling into a context tax. The two dials work against each other only if you neglect one of them.

The loop

1. Capture: when you correct the agent, it writes the correction down — what happened, what should have happened, which file.
2. Classify: one-off, recurring pattern, or broken rule.
3. Promote: recurring patterns go to memory or CLAUDE.md; rules the agent has broken go to a hook, because hooks run as code rather than as a request.
4. Review: the agent proposes the edit as a diff. You approve it like any other change.
5. Prune: on a cadence, the agent lists rules nothing has needed recently or that contradict the codebase, and proposes deletions.

Where it goes wrong

Self-approval. An agent that silently edits its own rules can relax its own constraints. Rule changes are diffs with a human signature, every time. Permissions and hooks protecting secrets or deploys should never be self-editable.

Rule inflation. Writing a rule for every hiccup is the same failure as never writing one, just slower. The promotion threshold — corrected twice — is the filter.

Vague rules. “Be careful with the database” is a mood. “Never run a migration without listing affected tables first” is testable. If you cannot tell whether a rule was followed, rewrite it.

What makes it better than a better prompt

Anthropic’s write-up on harnesses for long-running agents leans on persistent progress files and incremental, verifiable steps so each new session starts from recorded state rather than memory. Harness² applies the same discipline to the rules themselves: the lessons are recorded state, versioned in the repo, reviewed like code. Over time the harness becomes the most accurate document of how your project actually works.

A harness makes the agent reliable today. A harness for the harness is what keeps it reliable next quarter.