Nobody Wants to Read Your Pull Request Anymore

A ten-patch Linux kernel series written with LLM assistance by Linux creator Linus Torvalds delivered an order-of-magnitude speedup on a read-heavy benchmark by teaching folio lookup to detect hot folios and run data accesses inside an RCU-protected region without a fastpath refcount, but human review caught a race condition, leaving the series needing more work. A separate catalogue of how 94 prominent engineers run coding agents found only eight still read every line, with 31 using another model to review code and 21 spot-checking risky parts, while agents increasingly get their own version-control accounts and assisted-by provenance tags.

Three unrelated things landed this week, and together they point at the one resource agents do not make cheaper: somebody qualified to say yes. 1. Linux's Creator Shipped an LLM-Assisted Patch Series. Review Found the Race Condition. The Linux kernel just received a ten-patch series written with LLM assistance by the person who started the project https://lwn.net/Articles/1100449/ . The work is not cosmetic: it teaches folio lookup to notice when a folio is hot, delays releasing it until after an RCU grace period, then runs the data accesses — not just the lookup — inside the RCU-protected region, with no refcount on the fastpath when the user copy does not fault. The payoff is an order-of-magnitude speedup on a read-heavy benchmark. Review has already found a race condition, so the series needs more work. That last sentence is the whole story. This is the most scrutinized codebase in commercial computing, and the maintainer in question spent the spring complaining that AI had made bug reports free to produce and ruinous to receive. Now he is on the producing side, and the division of labor is what you would hope for: the machine moved the performance needle, human review caught the concurrency bug before a release. Notice where the difficulty went. The patch was the cheap part; establishing that a lockless fastpath is safe under every interleaving is the expensive part, and no tooling made that cheaper. The kernel's disclosure tag for assisted work admits as much — provenance matters because the reviewer has to calibrate how hard to look. Why it matters: - For ICs: The skill that appreciates is adversarial reading: concurrency, failure modes, invariants. Generation is table stakes; being the person who finds the race condition is not. - For leaders: If agents raised your team's output and review capacity is unchanged, you have not increased throughput. You have increased queue depth and moved risk downstream. - For founders: Hard systems work is suddenly within reach of small teams. The constraint is whether anyone on yours can validate it. - The general pattern: automation relocates work to whatever step it cannot do. Here that step is judgment under uncertainty, still staffed by humans. 2. Only Eight of Ninety-Four Prominent Engineers Still Read Every Line Someone assembled a catalogue of how 94 well-known engineers actually run coding agents https://www.prehistoriq.com/stacks/ , built from their public repos, rules files, talks and posts rather than a survey form. Most run more than one agent and 79 of the 94 changed their setup within the past year, but the review numbers are the interesting ones. Thirty-one have another model review the code before it lands. Twenty-one spot-check the risky parts and skim the rest. Seven rely on tests alone or ship with no review at all. Eight still read every line. What replaced reading is more inventive than "they stopped caring." One engineer runs two adversarial model reviewers whose job is to refute every change; several run cross-vendor review, one lab's model checking another's. One has the agent write an explainer plus a quiz on the change and merges only after passing it. Another reads a complexity-and-mutation-scored diagram instead of a diff. One triages each attempt for thirty seconds and rejects most before reviewing the survivor properly. Taken together, that is a shift from reading code to sampling it against a stated correctness condition. The second trend is provenance. Agents are being given their own version-control accounts so their commits are separable from a human's; model names are going into PR titles, assisted-by footers onto commits, exact versions into READMEs. The repeated principle is that whoever submits a change owns it, regardless of what typed it. That is the review bottleneck being managed socially rather than technically: if you cannot read everything, you at least want to know whose reputation is attached. Why it matters: - For ICs: Pick one verification habit and make it mechanical — a failing fixture before the fix, a second model's critique, a diagram you can scan. Good intentions do not survive a 2,000-line diff. - For leaders: Write down your review policy for agent output, because one is forming by default. Decide what gets disclosed too: "a human signed off" is not the claim "a human read it." - For founders: Review tooling is where the spend is going, and the buyers are the people who already gave up on reading diffs. 3. One Maintainer's Answer to AI Pull Requests: Turn Them Off The sharpest response to all this is to stop accepting pull requests entirely https://ghuntley.com/remix/ . The argument: the bazaar was already straining — maintainers nagged to merge, unpaid, owing an obligation they never signed — and machine-generated patches burst the dam. If a change was produced by an agent, reading it line by line buys the maintainer little — what is worth reading is the idea. Hence the proposal: publish source and intent under an as-is license, disable pull requests, keep discussions open, and document how to remix instead. The mechanism is what makes it more than a complaint. A consumer who wants different behavior writes a small Nix overlay expressing that change once, then reuses the derivation across dev shells, container images, system tests, CI and agent sandboxes without maintaining a fork. The worked example is patching Git to refuse force pushes locally rather than arguing for it upstream. Discovery moves to the fork graph, which an agent can summarize; maintainers sample ideas back when they feel like it. This is a repricing, not a philosophy. Licenses always said as-is; what made the patch the default unit of exchange was that remixing was too hard for most people, a knowledge gap agents have largely closed. The cost is real: a thousand private overlays is a thousand places a security fix does not reach, with no upstream to carry the advisory. The model buys maintainer sanity by pushing integration cost onto downstream users, and only works where a composition layer exists. Why it matters: - For ICs: Learn your stack's override mechanism. Getting a one-line behavior change without forking or waiting on a maintainer is becoming a normal skill. - For leaders: Audit how many dependencies you consume through local patches and who owns rebasing them. That burden grows quietly and comes due during an incident. - For founders: If you maintain an open repo, set contribution terms before the generated patches arrive. Silence reads as an obligation to review. The Verdict: Real or Hype? LLM-assisted patches in high-stakes systems code → Real. The output is good enough to be worth reviewing, and the review is still what makes it safe. Model-on-model review replacing line-by-line reading → Real but unevenly distributed. A third of the most sophisticated practitioners have moved; most organizations have not written the policy, let alone built the gates. Closing pull requests for a remix model → Real but early. Compelling for maintainers, unsolved for everyone downstream who needs a security fix to propagate.