AI-generated pull requests leave open-source maintainers rebuilding context

AI-generated pull requests are adding pressure to open-source maintainers because they often arrive without the context needed to assess them. CodeRabbit says the issue is less about code quality than whether agents and their operators understand a project's rules, existing work and contribution process.
- Merged GitHub pull requests rose from 25 million a month in January 2023 to 90 million in March 2026.
- Agents often miss repository rules and submit duplicate fixes.
- Projects are using AGENTS.md, CLAUDE.md and written AI policies to guide contributions.
- Human maintainers remain responsible for verifying and approving merges.
- Fatih Kadir Akın says context and verification are essential contributor skills.
More pull requests, less useful context
Merged pull requests on GitHub climbed from 25 million a month in January 2023 to 90 million a month in March 2026, with GitHub tying the increase to agentic workflows, according to CodeRabbit. As more changes reach open-source queues, maintainers are seeing submissions that ignore project instructions or duplicate fixes already proposed.
Fatih Kadir Akın, who built prompts.chat and works as a developer advocate on Automattic's WordPress team, described the problem in an interview with The Merge: “The code itself may be slop, or it can be really nice code. The problem is that whoever is controlling the agent isn't controlling it very well.” He said the missing piece is context, which maintainers must reconstruct during review.
At prompts.chat, a worker regenerates a CSV copy of the prompt database every day. Although the repository's pull request template and CONTRIBUTING.md tell contributors not to edit the file, agents continue submitting changes to it. Akın said, “the agents don't care.” He also described agents fixing the same issue multiple times without checking for an existing pull request.
Projects are putting instructions closer to the code
Open-source projects are responding by making their rules harder for agents to miss. AutoGPT placed instructions in AGENTS.md files beside the code they govern and linked to them from CLAUDE.md files. Superpowers added an agent-focused contribution guide to its CLAUDE.md and AGENTS.md files, with ignored instructions leading to automatic rejection of pull requests.
Home Assistant's written AI policy welcomes AI as an aid but closes pull requests or issues opened independently by an agent. Its project lead, Franck Nijhof, says AI makes contributions cheap to produce while maintainers still have to verify each one. GitHub has also introduced controls that let maintainers restrict pull requests to collaborators or limit how many a user without write access can keep open.
These measures address volume as well as quality. OpenCode was receiving around 200 new pull requests a day when Ryan Vogel appeared on The Merge. Daniel Stenberg has said curl cannot simply ignore AI-generated reports, because some identify real flaws, but the volume still creates a substantial review burden.
Human review and communication remain central
CodeRabbit reviews every prompts.chat pull request first, and GitHub Copilot sometimes takes a second pass, but Akın still reads the code before it is merged. “The biggest problem is that you have to be an approver as a human,” he said. “I'm merging it to something that works on the internet. I should be careful.”
Akın's advice for contributors is to treat context as part of the contribution. The agent needs enough information to produce a change that fits the project, while the maintainer needs an explanation that makes the change easier to trust. He also identifies verification as a core skill. Even when developers do not know a language themselves, understanding concepts such as memory leaks can help them recognize when generated code is wrong.
For users and maintainers, the source points to a hybrid workflow rather than fully autonomous merging. Clear instructions, limits on incoming work and human approval can help projects handle more agent-written contributions, while contributors still need to communicate why a change belongs and verify what the agent produced.
CodeRabbit: AI pull requests in open source have a context problem


