Google’s Sashiko system uses Gemini-powered agents to help review Linux kernel patches, handling about 1,200 reviews a day in production. Google says its use is associated with a 35 per cent drop in follow-up fixes within 60 days of merged commits, a promising result that is not proof the system caused the reduction.
Watch Desk analysis
What happened
Google’s AI and Infrastructure team describes Sashiko as an open-source agentic code-review system built for Linux maintainers. It separates finding potential bugs from checking them: parallel specialist sessions identify issues, followed by deeper, per-bug verification. The system has been used on kernel mailing lists for seven months, according to Google’s account of Sashiko.
Google reports that Sashiko completes around 1,200 reviews daily. It also says deployment correlates with 35 per cent fewer 60-day follow-up fixes for merged commits. That is an association reported by the team, not evidence on its own that Sashiko produced the change.
Why it matters
Kernel patches are scrutinised before they become part of software used across an enormous range of systems. Sashiko’s approach is notable not just for applying AI to that work, but for splitting bug detection from verification. In code review, a plausible-sounding false alarm is still work for a maintainer; checking findings is part of the job, not optional garnish.
The reported review volume and reduction in later fixes make this a useful real-world signal about AI-assisted review. The figures do not, by themselves, tell us how the system performs across different patch types or how much human review remains necessary.
Our read
The most encouraging detail is the verification stage. AI code review is only useful if it helps maintainers find real problems without burying them in confident nonsense. Sashiko’s reported results are worth attention, while the correlation deserves to stay a correlation until the evidence shows more.
For maintainers, the key question is whether the system’s findings are accurate and useful enough to save time, not whether it can produce a large daily review count. The kernel is a demanding proving ground, and that makes this more than another demo with a tidy interface.
What to watch
- Whether Google publishes further detail on how the 35 per cent figure was calculated.
- How often Sashiko’s findings are confirmed, dismissed or missed by maintainers.
- Whether the system’s approach spreads to other large, actively maintained codebases.
Discussion spark: For high-impact software such as the Linux kernel, should AI review be used to triage patches for maintainers, or is it ready to make substantive review decisions?
Sources and evidence
- Sashiko: Lessons from building an agentic code review system for Linux (18 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.