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
Criteria come from the review itself: no findings at or above a severity (e.g. medium) in chosen categories such as Security or Reliability, a small blast radius, and compliance with team rules.
Guards name repositories, paths (for example
auth/**or a payments service) and authors that are never auto-approved, including other bots and agents by their Git login. An author allowlist is also possible.Scope can be a whole workspace, a Git organization or one repository.
Audit: each approval names the workflow and threshold that authorized it, and a triage timeline shows which guard held a PR back.
Self-dismissal doesn't count: an author dismissing their own finding does not clear the way to approval.
The five questions to ask any auto-approver
Which changes qualify? Rules by finding severity and category, with stricter treatment for security.
Where can it act? Start narrow: one repo, with guards and author limits.
Can we observe first? Run in label-only mode and compare with human reviews before granting approval rights.
What happens after another commit? Approval must track the current head, not the reviewed one.
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
A one-line fix (session names breaking when a task description began with blank lines) got zero findings and was auto-approved; the PR was left unmerged. The author is explicit that this only shows the workflow runs, not that the review is accurate.
An earlier deliberately risky change, a debug endpoint for reading logs, drew seven findings, including anonymous log access, path traversal out of the log directory and server command execution. That is the kind of change the rules must stop.
The session-level review feature was not tested.
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
Throughput vs trust in the reviewer. Auto-approval is only as good as the review's recall. A missed finding becomes an approved PR.
Severity thresholds vs real risk. "No medium-or-higher findings" measures what the reviewer noticed, not what the change does. Blast radius and path guards are the backstop.
One approval vs required approvals. Counting the bot as one approval keeps code-owner rules intact. Raising it to "sufficient" removes the human from the loop.
Vendor lock-in. The policy lives in the vendor's workflow engine, not in your repo's branch protection.
Failure modes
Agent-on-agent approval. If agent authors aren't guarded, one agent's PR gets approved by another agent's review with no human ever looking.
Guard drift. A sensitive path renamed or added outside
auth/**silently loses its guard.Stale approval. Without re-review on push, a clean first commit carries approval for later changes.
Split-change blind spot. Two individually low-risk PRs that are risky together each qualify.
Small-sample confidence. A single clean approval says nothing about precision or recall.
Review checklist
Is there a shadow period (label only) with measured agreement against human reviews?
Are agent and bot authors guarded, or explicitly allowed with reasons?
Are guards defined by path and repo, and reviewed when the codebase moves?
Is approval withdrawn on every new commit?
Does the bot count as at most one approval, with code owners still required?
Can you answer, for any approved PR, which rule and which commit justified it?
Are multi-PR agent sessions reviewed as a unit before merge?
Related
R052 Open Code Review: Alibaba's hybrid AI code reviewer (the review stage these rules depend on).
R058 AI-native SDLC: Anthropic's playbook vs AWS AI-DLC (PR review as one play in a lifecycle).
R073 Twelve open-source repos for securing AI-written code (security-specialized reviewers).
R132 Fold subagents into the main loop (fewer, larger agent changes to review).