Rootly Abandons Small-PR Rule as AI Agents Reshape Code Review

Incident management platform Rootly has retired its long-standing practice of requiring small pull requests, challenging a development convention that served teams of human programmers but now impedes AI-driven workflows. The company argues that when machines write code, the traditional metrics for review—line count and diff size—no longer predict risk or catch real bugs.

Why Small PRs Made Sense Before#

For two years, Rootly enforced a disciplined small-PR culture where changes stayed within a few hundred lines, using stacked, atomic commits. This approach worked when humans manually wrote code; reviewers could absorb compact diffs and revert problematic changes quickly. The constraint reflected the pace of human typing and cognitive load during review.

AI Doesn't Think in Increments#

Agentic AI has inverted that equation. Rather than writing code line by line, AI agents generate complete implementations—migrations, data models, services, controllers, tests, and UI components in one pass. When Rootly tried forcing AI systems to produce stacked PRs instead, the results were technically sound but contextually broken: a reviewer's feedback on one PR hinged on architectural decisions made in another, creating review deadlock across multiple diffs.

The real problem AI introduces is not in syntax but in semantics. As Rootly's team notes, "AI bugs are context bugs. The code works, but it is applied to the wrong thing. A migration that drops a column still in use by a background job. A service that writes to a table another team reads from." Size became a useless signal; blast radius became the meaningful one.

Rethinking Code Review for Machines#

Rather than optimizing for faster human review of AI code, Rootly built an internal AI reviewer that evaluates every PR for risk, engineering standards, and the one question that matters: if this change breaks, what user-facing behavior fails? The tool flags impact-altering changes (what the system does) separately from optimization or cosmetic ones (how fast or how it looks), giving human reviewers a structured summary instead of drowning them in raw diffs.

This mirrors a broader industry shift. Rewind, a backup and versioning company, adopted a similar risk-based model in its AI code review tool, Diff Vader, concluding that PR risk has almost nothing to do with line count.