AI Agents Reshape Code Review: Rootly Abandons Small PR Rule

Incident management platform Rootly announced it is abandoning its long-standing 'small pull request' rule. Co-founder and CTO Quentin Rousseau explained that for the past two years, the company required stacked PRs and limited atomic changes to a few hundred lines of code. This made sense when humans wrote code, as small diffs are easier to review and roll back. But AI agents have changed the game: they think in terms of 'features' rather than 'increments,' and can output complete implementations in one go, including database migrations, models, services, controllers, tests, and frontend components.

Attempts to have AI generate stacked PRs failed: while the code had no technical errors, the overall business context was worse. When reviewing a PR, comments often depended on changes in another PR, forcing reviewers to switch between pages and increasing cognitive load. The team eventually realized that the small PR rule was designed for human coding efficiency. Once AI broke through human efficiency limits, the old rule became extra overhead.

Rootly's response was to build an internal AI code reviewer that examines each PR against engineering standards, generating a structured report with risk assessment, standardized scores, confidence scores, and issues categorized by severity. The key difference is that it doesn't act as a human reviewer but answers one question: if the change has defects, which user-facing features would break? The reviewer distinguishes changes that alter business behavior from those that only affect performance or UI, and assigns corresponding risk levels.

Rousseau emphasized that feature flags shift the safety boundary from 'merge' to 'release': every important feature is released under a feature flag, merged but disabled by default. The real review happens during progressive rollout: internal team, small customer subset, 10% of users, then all users. The team concluded: 'The size of the code change is no longer a meaningful metric; what truly matters is the blast radius of failures.'

Rootly's view is not isolated. Backup service provider Rewind stated that its code review tool Diff Vader borrowed Rootly's risk model, arguing that PR risk is almost unrelated to lines of code, and instead assigns PRs by risk tags. At QCon London 2026, Michael Webster discussed the impact of headless AI agents on delivery pipelines, noting that large-scale PRs generated by AI create bottlenecks for human review and accumulate technical debt.

At the AI-Native Developer Conference in London in June 2026, Patrick Debois and other experts explored how PR workflows become an anti-pattern at agent speed. Debois argued that PRs make sense in open-source communities because contributors need to build trust; but within teams with shared context, when agents iterate quickly, the PR review cycle is hard to justify. AI token consumption is quantifiable, and process inefficiencies directly show up in billing costs.

Rootly now focuses on issues that can predict production incidents. The 'why' and 'what' in a PR are filled in by the human using the agent, with explicit instructions that AI should not generate these, to capture context: why the change, why now, and the business need. Each PR must describe a safe rollback method, including necessary data fixes.

Rousseau admitted that abandoning the small PR rule was initially uncomfortable, but it was necessary to support 'rapid delivery of reliable software.' He elaborated on the shift to production-end safety in another article. The team concluded: in the era when all code was written by humans, the small PR model was optimal; but now that AI agents deliver complete features, that model no longer applies.