Tessl writes its own review standards into the repo: 75% of 1,200 weekly agent PRs merge without a human

Tessl's engineering team lets most agent-written pull requests merge without human intervention, turning review from a bottleneck into a process it can tune itself. The team previously used another AI reviewer whose findings were not real defects about a third of the time, so engineers reviewed the reviewer, checking each finding and each fix, and could not tune it to their own standards. In a one-week side-by-side run on one monorepo, the two reviewers raised roughly 1,850 findings; 91% of Tessl Code Review's findings were real defects against 68% for the previous reviewer, leaving about 435 fewer false positives to read and dismiss each week.

The reviewer checks every PR against the standards the team already writes for its agents, including coding conventions, design system and AI code review policy, so the team can see why a finding was raised and change what is checked. The previous AI reviewer applied the same general rules to every repo, so it could not tell when a change broke a rule specific to this codebase. On an external third-party benchmark with human-labeled defects the two reviewers tied; on this repo, with its own standards available, Tessl Code Review found more real defects and far fewer false positives, and led on all six change types the AI code review policy says always need a human check, with privacy at 27 versus 4 issues and access control at 94 versus 34.

The team turns patterns flagged repeatedly into verifiers, pass-or-fail checks that run on every change, closing the reviewer's gaps itself. With the previous AI reviewer, closing such a gap meant waiting on the vendor. On default settings each review round costs a flat 50 credits and the team's PRs average 2.2 rounds, so each PR costs about 35% less to review than before; changing the model, effort or number of review lenses bills the AI credits used, with a minimum of 50 credits a round.

The report addresses engineering teams running agents on their own repos whose review cost rises with PR volume. The demonstrated scope is a one-week side-by-side run over more than 600 PRs on Tessl's single monorepo plus an external third-party benchmark with human-labeled defects; a real defect is limited to something that would misbehave in production, such as wrong behavior, a broken contract, data loss, a security hole, a crash or a leak, while style, naming and preference do not count. The team says it will next mine code review data on a schedule and suggest updates to the context its agents use, so lessons from one review reach the next agent before it writes code, and will share results as they come in.

The authors state a trade-off: the previous AI reviewer was stronger on low-severity and cosmetic issues and caught some defects on this repo that Tessl Code Review missed, and the team switched because it can close each of those gaps itself by updating a review lens, a standard or a setting. The report also notes that 8.5% of PRs went through more than five review rounds, that those PRs used 47.4% of review loop spend, and that about one in seven rounds on those PRs was triggered by a merge conflict, two-thirds of which found nothing new. Several measurement details remain marked as placeholders, including the side-by-side dates, the start date for the roughly 75% figure, the weekly PR volume, and the period the 2.2-round average covers.