Eval Harness

Every prompt tweak in a long-running agent looks like an improvement in the moment. The only way to know is a graded run against fixed inputs. Loopkit already ships .claude/agents/verifier.md — that is your grader. Do not rebuild it.

The three-stage loop

inputs.jsonl  →  runner  →  outputs.jsonl  →  verifier (per row)  →  verdicts.jsonl  →  diff vs baseline

Each stage writes to disk. No stage holds the whole run in context.

Stage 1 — inputs.jsonl

One JSON object per row: {"id": "case-01", "input": "...", "expected": "..."}.

• 20-100 cases is enough for a signal. More is nice, not required.
• Include known-hard cases, edge cases, and a couple of trivial ones as sanity anchors.
• Freeze the file. Rev the eval with a suffix (inputs-v2.jsonl) when you change it. Never edit in place — you lose the baseline.

Stage 2 — runner

A dumb loop: for each row, call the model with the current prompt/skill, capture output, write {"id": ..., "output": ...} to outputs.jsonl. No grading here — just capture.

• Same temperature every run (usually 0 for evals).
• Same seed / model version.
• Log the git SHA of the prompt/skill under test in the file header.

If the runner is smart it will bias the eval. Keep it dumb.

Stage 3 — verifier

Fan out one subagent per row (see subagent-fanout). Each gets:

• The input.
• The expected output (or spec).
• The actual output.
• The verifier system prompt from .claude/agents/verifier.md.

Verifier returns strict JSON: {"pass": bool, "why": "..."}. Collect into verdicts.jsonl.

Diff vs baseline

Two runs of the same eval on two prompt versions → compare pass rates per case. What matters:

• Overall pass rate — the headline.
• Regressions — cases that were green and went red. These block ship.
• New passes — cases that were red and went green. These justify ship.
• Flappy cases — inconsistent across reruns. Investigate; may be genuine model nondeterminism or a bad case.

A change that raises the mean but adds regressions is usually a loss — the new failures are cases you already knew worked.

Red flags

• Grader is the same model that produced the output, with the same prompt. Self-grading is lenient. Use a different persona at minimum; ideally a different model tier.
• Eval passes 100% on day one. The cases are too easy, or the grader is a rubber stamp. Add adversarial cases.
• Eval takes >30 minutes. Fan out. A serial 100-case eval is a serial 100-case bottleneck.
• Grader sees your prompt under test. It will grade what you wanted, not what happened. Feed it only spec + input + output.
• Baseline lost. Without baseline, "improvement" is vibes. Commit verdicts.jsonl to git.

When NOT to do this

• One-off script — build a checklist, not a harness.
• Prompt that changes daily and won't stabilize — evals need a fixed target.
• Task where "correct" isn't checkable (open-ended creative writing) — use human eval or a rubric-based grader, not pass/fail.

The verifier is already yours. The harness is 100 lines of glue around it.