Almost half the developers at Anthropic's Code with Claude event in London on May 19 raised their hands when asked if they'd shipped a pull request in the past week written entirely by Claude. Most kept their hands up when engineer Jeremy Hadfield asked who had shipped that code without reading it at all. The moment captured how normalized AI-generated code has become inside a frontier lab where the product is now building itself.
Anthropic ran Code with Claude for the second year, timing the London kickoff to coincide with Google I/O in Palo Alto. The company has pushed Claude's coding abilities aggressively since Claude 4 launched last year. Claude 4.6, released in February, and Claude 4.7, released in April, have turned Claude Code into a tool developers trust to write, test, and ship production code with minimal human oversight.
The event drew developers from companies including Spotify, Delivery Hero, Monday.com, Lovable, and Base44, all of which are restructuring development workflows around Claude Code. Anthropic's explicit goal is to push automation as far as it will go. Instead of developers prompting Claude to generate code and then cleaning up mistakes, the company wants Claude to check and correct its own work without human intervention.
Key facts
- 01Almost half the developers at Anthropic's Code with Claude event in London shipped pull requests written entirely by Claude without reading the code.
- 02Anthropic released Claude 4.6 in February and Claude 4.7 in April, both of which significantly improved the model's coding capabilities.
- 03Claude Code now includes a 'dreaming' feature that consolidates notes from previous coding tasks to improve performance on specific code bases.
- 04Companies including Spotify, Delivery Hero, Monday.com, Lovable, and Base44 presented at the event on reshaping development teams around Claude Code.
- 05Katelyn Lesse, Claude engineering lead, said Claude is now 'probably as good as a midlevel engineer at writing code.'
Claude Code agents now write notes to themselves while working on specific tasks. When a subsequent agent works on the same code base, it reads those notes to learn from previous errors. A new feature called dreaming, announced two weeks ago, reads through all these notes and consolidates the information, spotting patterns and recurring issues. The system is designed to let Claude Code improve its performance on a particular code base over time without developer input.
“The default isn't 'I'm going to prompt Claude'—the default is now 'I'm going to have Claude prompt itself.'”— Boris Cherny, Head of Claude Code
Anthropic's pitch is simple: get out of Claude's way. Ravi Trivedi, an engineer at the company, described the philosophy as 'let it cook.' Human developers should not even see error messages when something breaks. Claude will handle test-and-tweak cycles until the code runs correctly. The workflow assumes Claude can manage the entire loop from generation to deployment.
The vision has critics. Developers on Reddit and Hacker News have raised concerns that AI coding tools pushed by managers chasing productivity gains are making software development harder, not easier. The complaint is that developers now spend more time reviewing generated code than they saved by not writing it. Security researchers have warned that AI-generated code can introduce vulnerabilities that slip past review.
Katelyn Lesse, Claude engineering lead at Anthropic, acknowledged the concern but said the issue is process, not technology. All of the old software development best practices still apply, she said, and some teams have lost sight of them in the rush to adopt AI tools. She added that some technical managers at Anthropic are exhausted by the volume of code their teams now produce. Claude is now as good as a midlevel engineer at writing code, Lesse said, though expert engineers are still needed for system design and harder troubleshooting.
Angela Jiang, Claude product lead, laid out the long-term target. The absolute end state is Claude being able to build itself. That requires Claude to design systems, write the code, test it, review it, and ship it without human input. Anthropic is working toward that capability, and the Code with Claude event was a demonstration that parts of that loop are already closed.
“I think the absolute end state we're trying to get to is Claude basically being able to build itself.”— Angela Jiang, Claude product lead
The shift is happening fast. Claude 4 could code, kind of, when it launched last year. Now, less than 18 months later, almost half the room at an Anthropic developer event is shipping unreviewed AI-written pull requests. OpenAI, Google, and Microsoft make similar claims about the volume of code their models write. The tools are getting better, and the guardrails are getting thinner.
The question is whether software development best practices can scale at the same speed as code generation. Lesse said they can, if teams stick to them. But the temptation to skip review and ship faster is strong, especially when the model is improving every few months. Anthropic is betting that Claude will close the gap between midlevel-engineer output and expert-level oversight before the accumulated technical debt catches up.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
