Code with Claude London: What Anthropic's First European Summit Revealed About AI-Native Development

Boris Cherny, the head of Claude Code, did not start his London keynote with slides. He started with a memory. In secondary school, he wrote TI-83 programs to automate his math homework and taught himself HTML so his Pokémon card eBay listings would sell faster. The code was simple, but the feeling was immediate: he had an idea, he wrote it down, and it ran. Programming grew more complicated after that — compilers, type systems, deployment pipelines — stretching the gap between a concept and its execution. At Code with Claude's first European stop, on May 19, 2026, Cherny argued that gap is closing again, and that the implications for how software teams actually work go well beyond convenience.

The First European Stop

The ExCeL London event drew engineers and product teams from across the UK and the continent for a full day of keynotes, workshops, and hands-on sessions. CPO Ami Vora, Cat Wu (head of product for Claude Code), Angela Jiang (head of Claude Platform product), Dianne Penn, and Katelyn Lesse joined Cherny on stage. Where the San Francisco summit in May was heavy on compute announcements and benchmark numbers, London's emphasis was on practice: how teams that have already embedded Claude Code into their build pipelines are reorganizing around it, and what the infrastructure picture looks like for organizations with strict data-residency requirements. Anthropic chose to hold a full European event rather than a restream partly because those requirements shape what is and is not possible for a large share of its enterprise customers.

Self-Prompting as the New Default

Cherny's central argument reframes what software development looks like when agents are the primary unit of work. In the current model, a developer writes a prompt, Claude produces code, the developer reviews and adjusts. In the model Cherny described as emerging, the loop is tighter and the human's position inside it is different.

"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, Code with Claude London keynote, May 2026

The practical meaning: Claude plans its own approach, executes, checks its output against defined success criteria, and revises — surfacing uncertainty rather than suppressing it. Writing a good initial prompt still matters, but the more consequential skill is defining outcomes precisely enough that Claude can evaluate its own work without a human reviewing every step. Teams at Spotify, Base44, and Legora all demonstrated workflows built around this pattern, covering tasks that had previously required active developer involvement at each decision point.

Infrastructure for Autonomous Pipelines

London was also the venue for two infrastructure additions aimed squarely at enterprise security and compliance concerns. Self-hosted sandboxes let companies run agent workloads on their own compute, keeping sensitive code and data off Anthropic's servers. MCP tunnels — an extension of the Model Context Protocol — allow those agents to reach internal databases and APIs without routing traffic through the public internet. Together, they address the two objections that enterprise security teams most consistently raise when evaluating Claude for production deployments. Combined with the Outcomes feature, which lets teams define success criteria that agents use to iterate and self-correct, the platform is now closer to something that can run multi-day tasks without a developer watching it.

A third addition, the Multiagent Orchestration layer, lets operators spin up fleets of specialized agents to break large tasks into parallel workstreams. The workshop sessions at London walked attendees through cases where a single complex refactor — say, migrating an authentication layer across a monolith — was split among agents working different modules simultaneously, with a coordinator agent aggregating their work and flagging conflicts.

What the Event Signals

The London event was less about what Claude Code can do on a benchmark and more about what a team of five or fifty actually does differently once Claude Code is embedded in their development loop. Anthropic's message, reinforced across sessions, is that the most important adaptation is not learning to write better prompts but learning to delegate more completely — and to catch problems at the system level rather than the line level.

The Dreaming feature, which lets agents recall and build on context from previous sessions, got extended workshop time in London. Instructors walked through multi-day agentic tasks where an agent working on a large codebase could resume after an overnight pause without losing the thread of what it had been analyzing — a practical requirement for any team using Claude on production-scale projects rather than isolated demos.

With Tokyo running today and an extended session for independent developers on June 11, Anthropic's conference circuit is global for the first time. Whether the "Claude prompts itself" framing settles into a durable engineering practice or becomes another demo-stage idea that stalls in real codebases will be answered not in London or Tokyo but in the pull requests that follow.