At Anthropic's Code with Claude event, developers are shipping pull requests they never read

Anthropic’s Code with Claude developer event in London on May 19 revealed a dramatic shift in professional software development: developers are now comfortable shipping pull requests written entirely by Claude without reading the code. According to MIT Technology Review, when Anthropic engineer Jeremy Hadfield asked attendees how many had shipped Claude-written pull requests in the last week, nearly half the room raised their hands—and when he followed up asking who had done so without reading the code, most hands remained up.

The Automation Pipeline Has Inverted

The timeline matters here. According to MIT Technology Review, Claude 4 released a year ago could code “kind of,” but the cascade of releases—Claude 4.6 in February and Claude 4.7 in April—has fundamentally changed the calculus. Anthropic staffers reported that at their own company, Claude now writes most of the codebase, including the code that powers Claude Code itself. OpenAI, Google, and Microsoft make similar claims about their own automated development pipelines.

Anthropic is no longer optimizing for “AI writes code, humans fix it.” Instead, according to Boris Cherny (who heads Claude Code) and engineer Ravi Trivedi’s presentations at the event, the new paradigm treats human code review as optional friction. The default workflow has shifted: developers now prompt Claude to prompt itself, with Claude testing, debugging, and iterating until code meets requirements—all without human intervention in the loop.

Self-Correcting Code Agents

A feature Anthropic announced two weeks prior to the event, called “dreaming,” extends this autonomy. According to MIT Technology Review, Claude Code agents now write notes to themselves, recording task-specific information that subsequent agents can reference. This inter-agent memory allows the system to accumulate context and refine its approach across multiple iterations without requiring a human developer to interpret error messages or guide the fixes.

The philosophical framing at the event was explicit: Ravi Trivedi presented the principle as “getting out of Claude’s way,” or colloquially, “let it cook”—a directive to let the agent work unsupervised to completion.

Why This Matters

This represents a structural break with how software development has worked for decades. If the trend holds, code review shifts from line-by-line human inspection to high-level acceptance testing—does the pull request solve the problem? Rather than verifying correctness through reading, teams may audit outcomes through integration tests, monitoring, and rollback speed. For organizations still structured around human code review, this creates immediate questions: How do you hire and retain developers if the core skill (writing code) is now automated? And for governance: who is accountable when a Claude-written pull request introduces a security flaw or performance regression that no human ever saw? Anthropic’s framing suggests the answer is “Claude’s test suite will catch it”—but that confidence has not yet been stress-tested at the scale of production systems outside Anthropic itself.