Skip to content

Sashiko: Google's agentic code review for the Linux kernel

Google published how its Gemini-powered, open-source reviewer works and what seven months of kernel mailing list data show.

Draftship Editorial4 min read

What Google published

Google's AI & Infrastructure team published a post dated November 19, 2026 on lessons from Sashiko, an open-source agentic code review system for the Linux kernel that runs on Gemini. Roman Gushchin and Eric Brewer wrote it. Google donated the project to the Linux Foundation earlier this year and runs it at sashiko.dev, reviewing patches on public kernel mailing lists at no cost to developers. The name comes from a Japanese stitching technique for reinforcing worn cloth. The full write-up is on the Google Cloud blog.

Why review is the bottleneck

Linux spans more than 35 million lines of code and absorbs over 85,000 changes a year. Google's framing is that development in large open-source projects has always been limited by review rather than code generation, so better coding agents do not speed things up on their own. More generated code piles up in front of the same small group of maintainers, including plausible-looking patches from people who do not understand the subsystem, plus AI-generated security reports that are often duplicates or wrong. Build bots and fuzzers catch crashes, not design mistakes.

How a review runs

Sashiko separates finding candidate bugs from verifying them, and each pass must record what it ruled out. It pulls in the code around the diff, such as modified functions, struct definitions and direct callers, plus the subsystem's domain guides, so changed lines are never judged alone. A planner picks which passes to run and sets a reasoning budget from patch complexity.

Up to seven specialist sessions then run in parallel, one per class of problem: commit intent, implementation match, control flow, resource lifetimes, locking, security and hardware. Each averages about 100 git tool calls and returns two lists, concerns it believes are real and concerns it ruled out, with supporting code. A lead verifier audits both: proven bugs pass through, proven dismissals are dropped, and everything else gets its own verification agent. That includes cases where a specialist dismissed a real bug by assuming an unseen caller or lock would prevent it. The rule is symmetrical, since neither a claim nor a dismissal is accepted without concrete code proof across all call sites. Reports are written to read like a maintainer asking questions on a mailing list.

Prompts started from open-source Linux review prompts begun by Chris Mason, with more than 100 contributors since. Sashiko recently switched to combining per-tree or per-version auto-generated notes on the code with subjective maintainer notes that cannot be derived from the code.

What the data shows

Over seven months, Sashiko completed roughly 1,200 reviews a day across more than 100 kernel mailing lists, emailing only lists whose maintainers opt in, over 70 of which have. More than 1,500 kernel developers replied on the lists, and pushback on false alarms fed back into the prompts and verification engine. Sashiko is cited in over 2,000 mainline kernel commits and in over 500 of this year's official kernel CVEs. Google cites LWN.net calling it an important part of the kernel process and the top bug reporter of issues in the 7.2 release cycle.

Across the 18 releases before Sashiko, Linux 6.2 through 6.19 and 254,000 commits, an average of 3.4% of merged commits needed a follow-up fix within 60 days. Since rollout that is 2.8% in Linux 7.1 and 2.2% in Linux 7.2, a 35% drop, with the 30-day rate falling from 2.3% to 1.4%. Across 21,500 tracked mainline commits, patches with a clean review needed a fix within 60 days 0.9% of the time, less than a third of the historical baseline. Patches merged despite a critical warning needed a fix 7.6% of the time, though the published text cuts off mid-sentence there.

Limits and open questions

Google says it is too early for strong claims and that other factors affect release-to-release numbers, so it does not credit Sashiko alone. Coverage is partial because only opted-in lists receive email. The post names the failure modes the design counters: models cutting corners to spend fewer tokens, models anchoring on one obvious issue, commit log text biasing results unpredictably or maliciously, and no persistent memory, which Google counts as an advantage because one patch can be reviewed against many branches.

What this suggests for people building with AI

This is inference. The transferable parts are process constraints rather than model size: split finding from verifying, make each pass state what it ruled out, and accept neither a claim nor a dismissal without evidence. That likely matters more than the model when you ask an AI to check its own work. Draftship's guide to writing effective prompts covers the asking side of that loop, which is where Draftship turns a plain-language description into a hosted app with a built-in database and sign-in.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.