Claude Code Is Writing Pull Requests That Developers Never Read

At Anthropic's Code with Claude event in London last week, a speaker paused mid-session to ask the room a question. Who had shipped a pull request written entirely by Claude Code in the past seven days? Nearly half the audience raised their hands. The follow-up was more pointed: who had shipped a Claude-written PR without reading the code first? Some hands stayed up.

That moment, reported by MIT Technology Review after the May 19-21 event, is an unusually clear datapoint about how quickly software development norms are shifting. The question of whether AI-generated code gets reviewed is not new. The novelty is that the answer came from a room full of professional developers, at a conference hosted by the company that built the tool, with no apparent hesitation.

What the London Event Revealed

Code with Claude London ran across two days, with keynotes and hands-on workshops anchored by Anthropic's engineering and product teams. Spotify and Delivery Hero were among the companies that took the stage to describe how they had restructured their development workflows around Claude Code over the past year. The sessions were not primarily about new model announcements. Anthropic did not ship a new model at the event. The emphasis was on agent infrastructure: coordination between coding agents, context-sharing across sessions, and tooling that would make multi-agent development faster and more reliable.

The unreviewed-PR moment cut through all of that. It was an informal survey, not a controlled study, but informal surveys taken in rooms of several hundred professional software engineers carry weight. The show of hands did not prove that unreviewed code is the norm. It proved that it is common enough to feel ordinary in that room on that day.

The Oversight Problem Anthropic Is Acknowledging

Katelyn Lesse, Claude Code's engineering lead, and Angela Jiang, its product lead, both addressed the concern during the event. The risk they described is not that Claude Code produces bad code more often than a human developer does. It is that the sheer pace of output is exceeding the rate at which developers can meaningfully review what is being merged. Code that ships without review is code that nobody fully understands, and that is a problem that compounds. Security vulnerabilities that escape a first pass tend to stay in production longer. Maintenance becomes harder when the people maintaining the code did not write it and have not read it.

Lesse noted that some technical managers at Anthropic itself are "exhausted" by the volume of code their teams now produce each week. That candor is notable. The company building and selling Claude Code is also grappling with the operational effects of Claude Code on its own internal engineering culture.

"A sudden flood of code generated and shipped without proper human oversight is kicking serious security and maintenance problems down the road." Angela Jiang, Claude Product Lead, as reported by MIT Technology Review, May 2026

The Tools Anthropic Is Building to Close the Gap

Anthropic's response is partly technical. Claude Code agents already write structured notes to themselves at the end of each session, recording what they learned about a codebase, what decisions they made, and what they left unfinished. When a different agent picks up the same project later, it reads those notes to get up to speed faster. The "dreaming" system goes further: a background process runs across accumulated notes, spots patterns in the same codebase, and synthesizes higher-level documentation that future agents can use without reading every session log. It is a form of institutional memory built into the tool itself.

CI integration is also in development. The goal is that routine code changes receive automated review coverage before they reach human eyes, so the oversight gap narrows rather than widens as adoption grows.

What Anthropic has not done, at least publicly, is slow the adoption of Claude Code or discourage teams from letting it write pull requests unsupervised. The tool is designed to ship code at pace. The engineering culture around it is still catching up. Whether tools like dreaming and automated CI review close the gap fast enough is a question the industry is now running a live experiment on, one raised hand at a time.

For more on the agent infrastructure announced at the Code with Claude events, see our coverage of the San Francisco conference. For the internal tensions this acceleration has created at Anthropic, see our earlier report on the Claude Code quality postmortem. And for a look at how Claude's model family is evolving alongside these developer tools, visit our models page.