AI writes the code now. Humans can't read it fast enough — and open source just proved it
More than 40% of GitHub pull requests were started by AI agents in September. ESLint paid $51,000 to contributors in three years — then gave up filtering the flood.
- GitHub reported that over 40% of pull requests on its platform were started by AI agents in September 2026, or 110 million agent-started pull requests. m:…4364
- ESLint retired its contributor pool programme after five years, having paid over $51,000 to outside contributors between 2022 and 2025, including $24,680 in 2025. m:…4364
- The Linux Foundation found 59% of surveyed maintainers reported growing contributions associated with generative AI, and maintainers estimated 25% of incoming pull requests were low-quality AI-generated submissions. m:…4364
- Coding agents produce changes faster than humans can inspect them, argues Larridin CTO Ameya Kanitkar, who says humans should spend more time reviewing the spec than the generated code. m:…3498
The flood
GitHub reported that over 40% of pull requests on its platform were started by AI agents in September 2026, a total of 110 million agent-started pull requests in a single month. That is not a trickle of automation helping out at the edges. It is the main flow of code into the world's largest open-source platform arriving from machines. m:…4364
The maintainers on the receiving end are feeling it. The Linux Foundation found 59% of surveyed maintainers reported growing contributions associated with generative AI, and the same maintainers estimated that 25% of incoming pull requests were low-quality AI-generated submissions. The cost is not in the bad code itself. It is in the hours spent reading it to find out it is bad. m:…4364
ESLint just put a number on that cost. The project retired its contributor pool programme after five years, having paid over $51,000 to outside contributors between 2022 and 2025, including $24,680 in 2025. The programme did not die because the money ran out — ESLint's decision to end it was not due to financial expense. When the review costs more than the contribution is worth, the rational move is to stop paying. m:…4364
The projects push back
Other projects are answering with rules instead of money. The Godot Foundation prohibits autonomous AI agents and substantial AI-generated code, and requires new contributors with three or fewer merged pull requests to obtain maintainer permission before proposing new features. LLVM prohibits using AI tools to complete issues labelled good first issue. Node.js added extra requirements for pull requests longer than 5,000 lines. m:…4364
The pressure is personal as well as procedural. Linux kernel developer Lorenzo Stoakes removed himself from several reviewer roles because of the volume of AI-generated patches. A reviewer quitting is a louder signal than a policy change: it says the work stopped being worth doing. m:…4364
Review the intent, not the diff
Larridin CTO Ameya Kanitkar argues the fix is to stop reviewing code the way we used to. Coding agents produce changes faster than humans can inspect them, so he says humans should spend more time reviewing the spec than the generated code. His analogy is the compiler: developers do not review the assembly code a compiler produces, they trust the compiler and review the source. Agents, in his view, deserve the same treatment. m:…3498
Shopify's Helix system is the working example Kanitkar points to. Agents handle implementation and verification, while humans stay accountable for checkpoints and adversarial reviews. Helix breaks work into small checkpoints that must pass tests and visual matching, each requiring two adversarial code reviews, and human approval is required by default. Shopify rebuilt its Shop app in 12 weeks under that system. The lesson is not that review disappears. It is that review moves up a level, from lines to intent. m:…3498
Render makes the same point from the other side, the side where code meets users. The platform defines an application as being in production when users or other systems rely on it, whether or not it has been formally deployed, and says production readiness is a set of practices to maintain rather than a status that is earned once. Agentic development makes it easier to deploy before verifying that an application can support its users reliably. The review bottleneck is not just a maintainer problem. It is a production problem, and it arrives the moment anyone depends on the code. m:…9562
What is still unresolved
- Whether GitHub's 40% agent-started share keeps rising, and what share of those pull requests ever get merged.
- Whether spec-level review catches the security flaws that line-level review used to catch.
- Whether paid contributor programmes return in a different form, or stay dead.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.