Linus Torvalds, Sashiko, and the Reality of Agentic Engineering

With Linus Torvalds putting his foot down on AI in the Linux kernel, we look at the design of the Sashiko patch reviewer and why local-first agent workflows are the only way to manage the 'false positive' tax.

The developer world woke up today to a classic, uncompromising Linus Torvalds email on lore.kernel.org.

Responding to a heated debate on linking the Linux kernel's Patchwork tracking system with Sashiko—an agentic LLM code review tool—Linus drew a line in the sand:

"Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away. AI is a tool, just like other tools we use. And it's clearly a useful one."

This isn't just drama; it is a major architectural debate about how human engineers and AI agents collaborate on highly complex systems.

Here is what the Sashiko debate means for the future of development, and why it proves that the "agent-as-a-tool" model—the very core of Rope Notes—is the only sustainable way forward.

What is Sashiko?

Created by Roman Gushchin and built on prompt strategies pioneered by Chris Mason, Sashiko (named after the Japanese decorative reinforcement stitching technique) is an agentic review daemon. It monitors kernel mailing lists, pulls proposed patches, runs multi-stage LLM evaluation protocols, and points out bugs before human maintainers even open the email.

According to its maintainers, Sashiko catches around 53% of bugs that would otherwise slip completely past human reviewers. It is highly effective, and as Linus pointed out, "it keeps finding embarrassing bugs."

The Core Conflict: The False-Positive "Tax"

The dispute that triggered Linus's intervention wasn't actually about whether AI is smart. It was a logistical bottleneck.

Kernel developer Laurent Pinchart argued that if maintainers use Sashiko to review patches, those maintainers must fully verify and filter the AI's feedback before forwarding it to patch authors. The concern is obvious: automated tools that spam human developers with false-positive complaints or context-blind critique of unchanged code are a major tax on developer sanity.

But Roman Gushchin rightly pointed out that forcing maintainers to pre-triage every single AI warning defeats the purpose of automation. It just shifts the bottleneck.

Why the "Chatbot in a Web Browser" is Broken

The kernel debate highlights exactly why the industry's current approach to AI integration—clunky browser windows, copy-pasting code back and forth, or fire-and-forget autonomous bots—is fundamentally flawed.

If an agent is completely decoupled from your local environment:

  • It lacks the precision compiler, Git, and structural context of your local tree.
  • It creates a disconnect where "the AI proposes" but you have to manually clean up its mess.
  • The developer is treated as a passive supervisor sorting through a wall of text.

The Rope Notes Philosophy: You Own the Rope

This is exactly why we built Rope Notes. We don't believe in hands-off automated agents that dump code or reviews into your lap and run away.

We believe the developer must stay intimately connected to the editing loop:

  • Local-First Verification: Rope Notes keeps the LLM agent close to home. Because the Rust-based rope core and Dart analysis live right on your machine, your agent isn't guessing; it is referencing your exact files and syntax trees.
  • Interactive Git Diff Previews: When our agent operates in Execute mode, it doesn't just overwrite files or throw comments at you. It generates clean, inline ghost previews and Git diffs. You accept or reject changes interactively—no context-switching required.
  • Permissions by Default: You decide exactly what the agent can touch via .rope_notes/permissions.toml. You retain absolute architectural control.

As Linus wrote: "AI is a tool, just like other tools we use." It isn't a replacement for engineering judgment. By bringing agentic execution directly into a high-performance, offline-first editor, Rope Notes lets you wield that tool safely and at blistering speed.

Whether you are hooking up a massive cloud frontier model like the new Kimi K3 or running local Ollama weights on your machine, Rope Notes ensures that the agent works with you, not against your inbox.

Ready to build with a real engineering agent? Download Rope Notes today.

Further Reading & Resources