R141 · AI-built software and its verification › AI code review

Auto-approving agent PRs: policy-gated approval, guards and session review

Read the original

code-reviewgovernancesoftware-factoryhuman-in-the-loopvendor

Sources: Elvis Saravia, Auto-Approving PRs for Coding Agents (2026-11-18, written in partnership with Qodo), and Qodo's launch post, Introducing Auto Approval Workflows and Agent Session Review (2026-11-18). The feature is a vendor beta; the hands-on test is small and sponsored. Read this for the design pattern, not as an independent evaluation.

The problem

Once coding agents write code, run checks and open PRs, the queue moves to approval. Dependency bumps, test-only changes and copy fixes across many repos sit waiting for a human click. Auto-approval lets PRs that meet explicit rules proceed, so people spend review time on the changes that need it.

How the pattern works

PR opened (often by an agent)
  -> AI review: findings with severity + category, blast-radius estimate, rule compliance
  -> workflow: IF criteria met AND no guard matches
        THEN approve as bot (counts as ONE approval; code owners still apply)
          or add a "bot approved" label only (shadow mode)
  -> new commit? re-review; withdraw approval if it no longer qualifies
  -> human still merges

The five questions to ask any auto-approver

  1. Which changes qualify? Rules by finding severity and category, with stricter treatment for security.

  2. Where can it act? Start narrow: one repo, with guards and author limits.

  3. Can we observe first? Run in label-only mode and compare with human reviews before granting approval rights.

  4. What happens after another commit? Approval must track the current head, not the reviewed one.

  5. Can we explain an approval later? Every decision traceable to a commit, a rule and a workflow.

These apply whichever tool you use.

What the hands-on test showed

Session review: the cross-PR gap

Agent-built features often span several PRs and repositories. Each diff can look fine while the combination breaks compatibility or deploy order, or quietly does more than the plan said (the vendor's example: re-review triggered on a settings change nobody asked for). The proposed answer is a per-session "work package": deviations from the plan, merge order and interaction points, a reading plan, and risks tied to evidence. It is a research preview, so treat it as a direction, not a result.

Tradeoffs

Failure modes

Review checklist

Connections

See R141 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.