Linear’s review habit: recover the customer problem before reading agentic diffs

Primary account: Maciek Pekala, “Reviewing code in the agent era,” Linear, June 10, 2026. An engineer describes his own review sequence; this is practice testimony from a tool vendor, not a measured evaluation of defect detection.

What changed. Pekala reports issues created and PRs merged both increased 50% at Linear between January and March 2026; he felt pressure to rubber-stamp. His current sequence: review inbox; retrieve motivating customer feedback/bug/Slack thread; use a preview build for visible changes and share screenshots/recordings; scan scope and use grouped guided diff for sizable changes; ask whether the change solves the right problem and fits the system. He says agents cover much bug and logic checking, so he spends less time on line-by-line reading. When he sees a change solving the wrong problem, he draws the team back into discussion; minor revisions remain inline, with human or agent editing. He sees review as a rare peer-to-peer cultural junction while people otherwise work alone with agents.

What it does not prove. “Quality is holding up” is his confidence, not independently tracked quality; no human minutes saved, missed-defect rate, incident rate or post-merge maintenance numbers. This is a human-review approach, not evidence that such a gate can be removed. The 50% is a team before/after issue/PR count, not a causal agent-productivity estimate. Connect to review-autonomy question, feature-to-outcome and Linear’s support-to-code workflow.