Artificial intelligence news.
Straight from the original announcements.

Google details Sashiko, its agentic code reviewer for the Linux kernel

November 18, 2026, 18:00By the AI News Bites Editorial TeamSource: Google Cloud
Google details Sashiko, its agentic code reviewer for the Linux kernel
Image: Google Cloud

Google has published lessons from building Sashiko, an open-source agentic code review system powered by Gemini for the Linux kernel. The system is designed to help maintainers identify bugs in proposed patches while reducing the burden on the small group of experts who review them.

Key points
  • Sashiko is an open-source code review system powered by Gemini.
  • It uses up to seven parallel specialist sessions for each review.
  • The system completes roughly 1200 reviews a day across over 100 kernel mailing lists.
  • More than 70 mailing lists have opted in to receive its review emails.
  • Sashiko is cited in over 2000 mainline kernel commits and over 500 official Linux kernel CVEs.

Why Linux code review is a bottleneck

Google says development in large open-source projects has traditionally been limited by code review rather than code generation. Faster coding agents can produce more patches, but without a way to establish trust in those changes, maintainers still have to assess them by hand.

The challenge is particularly significant for Linux. The kernel spans more than 35 million lines of code, and over 85,000 changes land every year. It runs billions of Android phones, most cloud providers and nearly every supercomputer. Maintainers review hundreds of proposed patches each day to find issues such as memory leaks, deadlocks, race conditions and security bugs.

Google says automated build bots and fuzzers such as Syzkaller can detect build errors and crashes, but code review remains central to checking the stability and safety of the kernel. The company built Sashiko to help maintainers handle this workload and catch bugs before patches are merged.

A multi-stage review process

Sashiko separates the discovery of possible bugs from their verification. It first gathers context around a patch, including modified functions, structure definitions and direct callers, and loads domain guides for the relevant subsystem. A planner then selects the specialist reviews needed and sets a reasoning budget based on the patch's complexity.

Up to seven specialist sessions run in parallel. Each focuses on one area, including commit intent, implementation match, control flow, resource lifetimes, locking, security or hardware. The sessions use Git tools to search the repository, averaging about 100 tool calls per review. They record both suspected problems and candidate concerns they investigated and ruled out.

A lead verifier then triages those results. Difficult cases receive separate verification agents, each assigned one specific question. Sashiko requires concrete code proof across all call sites before accepting either a bug claim or a dismissal. It then produces a report written in the style of a human maintainer asking questions on a Linux mailing list.

Sashiko's reported use in the kernel community

Google says Sashiko reviews public patches at no cost to developers and runs the service at sashiko.dev. The project was donated to the Linux Foundation earlier this year. Its prompts and operation are open, including every prompt, tool call, reasoning step and ruled-out bug. More than 100 contributors have helped improve the prompts and the system, including kernel maintainers and engineers.

After seven months of production data from kernel mailing lists, Sashiko completes roughly 1200 reviews a day across over 100 lists. It sends emails only to lists whose maintainers opt in, and more than 70 have done so. More than 1500 kernel developers have replied to its reports, with feedback on false alarms contributing to improvements in the prompts and verification engine.

Google says Sashiko is cited in over 2000 mainline kernel commits and in over 500 of this year's official Linux kernel CVEs. According to the company, the system can help prevent bugs before a patch is merged, while also uncovering preexisting issues during review. That positions the project as an AI-assisted layer for the kernel's existing maintainer workflow, rather than a replacement for human review.

Related news

Abundant runs more than two million AI rollouts on Modal each month
For developersNovember 19, 2026

Abundant runs more than two million AI rollouts on Modal each month

Abundant has built its reinforcement learning rollout platform primarily on Modal, running more than two million Sandboxes per month as of November 2026. The company…

GitLab says Duo Agent Platform credits will go up to twice as far
Products and toolsNovember 19, 2026

GitLab says Duo Agent Platform credits will go up to twice as far

GitLab says its Duo Agent Platform now uses fewer model requests for agent and flow tasks, allowing most customers’ GitLab Credits to go up to twice as far. The new…

Vercel adds Google’s Gemini 4 Neon to AI Gateway
ModelsNovember 18, 2026

Vercel adds Google’s Gemini 4 Neon to AI Gateway

Vercel has made Google’s Gemini 4 Neon available on AI Gateway, adding a model aimed at coding, agent workflows and multi-step tasks. The model supports multimodal…