Incident management platform provider Rootly has published an account of its decision to drop its long-standing small pull request rule, arguing that the practice no longer serves its purpose now that AI agents generate most of its code. The company describes a shift from measuring PR size to assessing blast radius, with feature flags and rollback capability taking precedence over line counts.
In the post, Rootly co-founder and CTO Quentin Rousseau explains that for two years, it enforced a strict small-PR culture with stacked PRs and atomic changes limited to a few hundred lines. He states that this made sense when humans wrote code by hand, as smaller diffs were easier to review and revert. He writes that AI agents have changed the equation because they think in features rather than increments, producing complete implementations, including migrations, models, services, controllers, tests, and frontend components, in a single output.
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.
- Rootly engineering team
The post describes how attempts to make AI agents produce stacked PRs resulted in technically correct but contextually worse outcomes. Review comments on one PR often depended on decisions made in another, forcing reviewers into mental gymnastics across multiple tabs. They figured out that the small-PR rule was optimised for human writing speed, and AI removed that constraint while turning the rule into overhead.
Rootly's response was to stop reviewing AI code the way it reviewed human code. The company built an internal AI code reviewer that reviews every PR against engineering standards and produces a structured review with a risk assessment, standardisation score, confidence score and specific findings grouped by severity. The key difference is that it is not trying to be a human reviewer; instead, it answers one question per PR: if this change has a bug, what user-facing behaviour breaks?
The AI reviewer flags changes that alter what the system does, versus changes that affect how fast or how something looks, and assigns appropriate risk profiles to each, thus giving the human reviewer a structured starting point instead of just a raw diff.
Rousseau emphasises how the use of feature flags has moved the safety boundary from merge to rollout, as every significant feature now ships behind a flag. When the PR has been merged and the code pushed to production, it’s turned off. The real review happens during progressive rollout: with the feature enabled for the team first, then a handful of customers, then 10 percent, then everyone
The size of the diff stopped being the useful signal. The blast radius is.
Rootly engineering team
In a presentation at QCon London 2026, Michael Webster discussed the rise of headless AI agents and their impact on software delivery pipelines. He shares how massive, AI-generated pull requests create a severe bottleneck for human reviewers and introduce persistent technical debt. Others in the industry are thinking similarly; Rewind, a backup and versioning provider, recently described how it adopted Rootly's risk-based model for its AI code review tool, Diff Vader. In a post on its blog, the Rewind team wrote that a PR's risk has almost nothing to do with its line count. Diff Vader assigns a risk label to every PR based on the review's findings, not on how many lines changed.
Pull requests at agent speed is a talking point at industry events. At AI Native DevCon London in June 2026, a panel featuring Patrick Debois, now DevRel at Tessl and widely credited as the godfather of DevOps, discussed why PR-based workflows are an anti-pattern inside enterprises once you are moving at agent speed. Debois argued that pull requests make sense in open source, where contributors might not be strategically aligned and trust has to be earned, but inside a team that shares context and goals, the overhead of a PR review cycle is increasingly hard to justify when agents are moving fast. In a podcast episode recorded at the event, Debois and fellow panellists described how agent costs are forcing process discipline. With humans alone, it was hard to see inefficiencies, but now, with measurable token costs, waste literally shows up in the bill.
Rootly's philosophy is now to ask the questions that actually predict production incidents. The 'Why and What' sections force the author to explain the motivation and scope of a change as well as its likely impact. For AI-authored PRs, the human who prompted the agent fills these in. Rootly explicitly instructs AI assistants not to generate these sections, because the whole point is capturing context the AI does not have: why this change, why now, what is the business reason. Every PR needs to describe how to safely undo itself, including any data fixes.
Rousseau closes out the post by explaining that killing the small-PR rule - a process that felt virtuous - felt uncomfortable at first, but was necessary to support the outcome of shipping reliable software quickly. He wrote more about this shift toward production-side safety in a separate post titled Stop Trying to Review AI's Code Faster: Bet on Rollbacks Instead.
In conclusion, the team write that small PRs were the right answer for a team of humans writing code by hand, but not for a team orchestrating AI agents that ship complete features.