Rootly Drops Small PR Rule as Agentic AI Changes Code Review Economics

Rootly’s engineering leadership observed that AI agents now emit complete feature implementations, including migrations, services, and tests, in a single pull request, making the historic rule limiting PR size irrelevant. The rule had been useful when developers wrote code manually, because smaller diffs were easier to review and revert. With AI, the constraint became a bottleneck, as reviewers had to juggle interdependent changes spread across many stacked PRs. Consequently, Rootly announced the removal of the small-PR rule and a shift toward assessing the potential impact of a change rather than its line count.

To replace the lost safety net, Rootly built an internal AI code reviewer that automatically evaluates every PR against engineering standards and emits a structured risk assessment. The reviewer produces a confidence score, a standardisation rating, and a severity-ordered list of findings, answering a single question: if the change is buggy, what user-facing behavior breaks? This approach treats the AI reviewer as a tool that surfaces risk, not as a surrogate for a human reviewer, allowing engineers to focus on the most critical issues identified by the model.

Adopting this new workflow requires integrating the AI reviewer into the CI/CD pipeline, maintaining its rule set, and training engineers to interpret risk scores and severity tags. The organization also had to adopt a feature-flag strategy, ensuring that any significant change is merged behind a flag and only exposed during a staged rollout. The upfront cost is the development and upkeep of the AI reviewer plus the operational discipline of flag-driven releases, but it eliminates the time spent on manual diff triage for large AI-generated PRs.

The new model stops being effective in scenarios where the AI reviewer cannot reliably detect context-specific bugs, such as migrations that affect background jobs still in use, or when feature flags cannot be applied (e.g., low-level library changes). In those cases, traditional human review or additional testing safeguards remain necessary. Moreover, teams that lack mature flagging infrastructure may find the risk-based approach harder to adopt without substantial process changes.

Rootly’s experience is already influencing peers; another company, Rewind, adopted a similar risk-labeling tool called Diff Vader, confirming that line count is no longer a useful proxy for risk in AI-driven development. Industry panels are debating whether PR-centric workflows are an anti-pattern at “agent speed,” suggesting that the shift may become a broader industry norm rather than an isolated experiment.