Agent review harness: every defect is also a harness defect
Sep 30 nick-delgado: review-agent-pr → address-pr-review → improve-agent-process — parallel standards/smell/spec/test reviewers, falsifying verifier, root-cause log on a tracking issue that batches process PRs.
Source: agent-review-harness — review AI-authored PRs and improve the authoring agent’s process
nick-delgado/agent-review-harness (skills for Claude Code / Codex / Antigravity; GitHub via gh; repo from Sep 30, still pushing) treats an ordinary PR review as incomplete. Listing defects is table stakes; each defect is also evidence about the authoring agent’s setup — instruction files, skills, specs, guardrails — and recurring causes should become concrete process changes before the next flood of agent PRs.
Three skills close the loop. review-agent-pr (fresh reviewer session) runs a preflight → intake → spec find → context manifest, then four parallel reviewers in fresh context each (standards, code smells / agent-typical patterns, spec alignment, test adequacy), a verifier that tries to refute every finding, and a root-cause analyst. It posts one PR comment (findings only, grouped Fix now / Needs owner decision / For the owner, with file:line and quoted rule/spec) and logs causes + proposals on a tracking issue labeled agent-process. address-pr-review is the authoring agent’s job: fix the “Fix now” set and reply finding-by-finding with start/end commits. After several reviews, improve-agent-process reads the tracking issue and opens one batched PR of docs, skills, tests, and CI checks so the same defect class does not return. CI owns tests/linters/builds; the harness reads results and does not re-run them. Audit trail is append-only comments — earlier reports are never edited.
Honest scope: GitHub-only today; zero stars at writing; process improvement quality depends on humans actually merging the batched harness PRs. The scarce idea is the feedback object: review debt becomes harness inventory.
Why it matters for code review
Verification tax under sustained agent volume is not only “more eyes on diffs” — it is a loop that spends review findings on the factory that produced them. Continuous harnessed momentum compounds when morning triage updates skills and gates, not just patches the latest PR. Review agent-authored merges for Delgado-shaped receipts: falsified findings, cause logs, and a process PR that closes the class. Pair with Roborev’s commit-time review loop when findings must re-enter the worker while context is warm, and with verify-first’s rejection tax when in-loop gates beat post-merge archaeology.