OpenHands verification stack: critic before push, review+QA on the PR, iterate until green
May 6 OpenHands: agent-level critic gates broken trajectories; repo-level review skill + behavioral QA; /iterate closes the loop — 1,000+ reviews, 58% faster MTTM, ~25% coverage lift on their SDK repo.
Source: The Verification Stack
OpenHands’ May 6 post names the bottleneck continuous factories keep rediscovering: agents made generating code cheap; verification — correctness, conventions, merge trust — is what saturates. Their answer is a layered stack that catches different mistake classes at different stages, so humans only spend time on changes already worth merging.
Layer 1 is an agent-level critic: a small, fast model scores the trajectory while the agent is still working, before any push. Low scores stop the run early or trigger retry, so broken attempts never become review load (trained and measured in their earlier Learning to Verify post). Layer 2 is repo-level: a code-review skill on every PR (structured checklist prioritizing data structures, security/correctness, real-path tests over mock-only, supply-chain risk; risk verdict 🟢/🟡/🔴; customizable via .agents/skills/custom-codereview-guide.md read from the PR branch) plus a QA agent that runs the software — Understand → Setup → Exercise → Report — with evidence (commands, outputs, screenshots) rather than diff reading alone. Closing the loop is the /iterate skill: open a draft PR, poll verification layers, fix, push, repeat until green, then mark ready — pairable with org-wide Automations so mechanical babysitting finishes before a human looks.
Production receipts from six months on OpenHands/software-agent-sdk: 1,000+ reviews across 900+ PRs and 38 releases; ~90% of PRs touch the automated review path; mean time to merge down 58%; test coverage up nearly 25%; bot precision approaching human while humans still review after the bot ~85% of the time. Static signals (complexity, smells, Bandit findings) improved even as the repo grew 144% — the maintainability-index dip is mostly LOC counting against itself.
Why it matters for code review
Continuous verified throughput is not “more Claude overnight” — it is an unskippable judge inside the generation loop and again at the PR. Review this stack as merge infrastructure: trajectory critic before push, independent review+behavioral QA on the diff, iterate until the mechanical gate is green. That is how always-on factories compound past burst mastery without turning morning review into a regression detector — and how FOMO for idle agents converts into inventory you can actually merge.