Contents
Merging pull requests without reading the code requires reliable quality checks and explicit agreement that passing those checks authorizes a merge. If developers are blamed for not reading the code even when they follow the agreed process, human review remains an implicit requirement.
Key Points#
- Automated checks must verify requirements, confirm that review findings have been resolved, and block merging when required evidence is missing.
- The team must agree on which changes can proceed without human approval and who is responsible when defects occur.
- Design discussions and knowledge sharing need a place outside mandatory pull request reviews.
- Start with a narrow scope, define escalation rules, and audit the results.
Define Merge Criteria#
“No-look merging” means merging changes without a person reading the code. Even so, evidence that the changes meet the team’s quality standards is still required.
Express those standards as explicit conditions. These four are a useful starting point.
1. Verify the required behavior with tests, including relevant failure cases. 2. Check applicable design and security rules through automated review. 3. Resolve review findings and recheck the revised code. 4. Ensure no required checks are skipped, inconclusive, or unresolved.
A passing status is useful only when the checks behind it cover significant risks. Before removing human approval, examine what those checks actually verify. Unless requirements are translated into meaningful tests, passing tests does not prove that the requirements have been implemented.
Make Approval an Explicit Policy#
Introducing AI review does not automatically delegate merge authority. Some teams accept AI-generated feedback while still expecting developers to make the final decision.
Clarify this distinction before changing the workflow. The engineering lead and relevant team members need to agree on eligible changes, the evidence that authorizes merging, and the conditions that require human judgment. Record that agreement somewhere contributors can find it.
Accountability must follow the same policy. If a defect slips through despite a developer following the approved process, investigate where the process failed. Assign someone to handle the incident and improve the checks. Blaming developers solely for not reading the code silently reinstates mandatory human review.
Preserve the Other Roles of Review#
Code review often serves several purposes at once: checking implementation quality, resolving design questions, and helping colleagues understand changes. Automating the first does not automatically fulfill the others.
Give these activities an explicit place in the workflow.
- Discuss consequential design decisions before implementation.
- Share behavior changes through documentation or demos.
- Record the architectural constraints that automated reviewers should check.
- Define which design decisions AI agents may make independently.
This prevents a practical problem: removing review approval while still relying on review threads to understand the system. Teams still need ways to question decisions and exchange feedback, even when those discussions do not block every merge.
Automate the Agreed Workflow#
Orchestration tools can enforce the process approved by the team. TAKT, for example, lets you define execution order and branches based on results: implement and review a change, return to fixes if issues arise, and proceed once required checks pass.
The distinction here matters. Deterministic control of a workflow does not guarantee that an AI reviewer’s judgment is correct. Reliably executing a review step does not prevent that review from missing defects. Evaluate both whether the process runs as intended and the quality of its conclusions.
Make stopping conditions explicit. If a required check cannot run, progress must stop. The same applies when review results conflict or a change falls outside the approved scope. Escalate these cases to the designated person or role with enough information to make a decision.
Start Narrow and Inspect the Results#
Do not remove human approval from every pull request at once. Choose one narrowly scoped category, such as documentation updates or interface changes within an existing component. Confirm that checks can distinguish eligible changes from those requiring escalation.
For the initial trial, sample 10 merged pull requests for inspection and examine every reported defect. Check whether tests covered the required behavior, review findings were resolved, and escalation rules worked. Ten is a practical starting sample, not proof that the process is safe.
Revise the policy based on what you learn. Missed defects may call for better tests, stronger review instructions, or a narrower scope. Repeated integration failures may require examining dependencies or directory structure. Expand the scope only when the evidence supports it.
Next Steps#
Write a one-page merge policy for one category of change, specifying required checks, escalation conditions, and who is responsible when defects occur. Get agreement from the engineering lead and relevant contributors before enabling automatic merging.
Source: https://zenn.dev/nrs/articles/c777388530aa84
Contents
