Anthropic and Claude: The AI Redefining Programming in London

The event Code with Claude, organized by Anthropic, took place in London over two days, starting on May 19, coinciding with Google's I/O conference in Palo Alto. Anthropic employees emphasized that this was merely a coincidence. Jeremy Hadfield, an engineer at Anthropic, spoke to the audience about using Claude to submit pull requests—those corrections or software updates submitted for review. Nearly half of the attendees raised their hands, indicating they had used Claude to fully draft these requests.

Hadfield then asked who had submitted a pull request entirely written by Claude without having read the code at all. This question elicited nervous laughter from the audience, but most hands remained raised, illustrating the growing confidence in these AI tools.

Pull requests represent the heart of software development, consisting of the pieces of code that professional developers spend their time writing. However, with the advent of tools powered by large language models (LLMs) like Anthropic's Claude Code and OpenAI's Codex, the way software is created has been profoundly transformed. Major tech companies, including Anthropic, now boast about the amount of code generated by these tools. Hadfield stated that "most of the software at Anthropic is now written by Claude," highlighting the extent of this change.

This rapid transformation has become the norm, and the London event was just one of many organized by Anthropic, with similar editions in San Francisco and Tokyo. This was the second consecutive year that Anthropic held developer events. Around the same time last year, the company had just launched Claude 4, which could code to some extent. However, with recent updates, including Claude 4.6 and 4.7 released in February and April, Claude Code has become a tool that more and more developers are adopting.

Anthropic aims to push automation to its limits. The goal is for Claude to check and correct its own work, thereby reducing human intervention. Boris Cherny, head of Claude Code, explained that the default process is now to let Claude self-evaluate. "The default is not ‘I will ask Claude’—the default is now ‘I will let Claude ask itself’," he stated during the opening conference.

If all goes well, human developers shouldn't even see error messages when something goes wrong. All of this will be managed by Claude, who will test and adjust, test and adjust, until everything works as it should. As Ravi Trivedi, an engineer at Anthropic, said during another presentation: "The key principle is to let Claude do it. We like to say: ‘Let it cook.’"

Trivedi introduced a new feature in Claude Code, announced two weeks ago, that the company calls dreaming. Claude Code agents write notes for themselves, recording and saving useful information about specific tasks. When another coding agent then starts working on the same code, it can use the notes to get up to speed more quickly and learn from mistakes that previous agents may have made. Dreaming is a system that Claude Code uses to read all these notes and consolidate the information they contain, spotting patterns and common issues across different tasks. In theory, dreaming should help Claude Code learn about a particular codebase and improve increasingly in its work.

The event also showcased success stories from companies like Spotify and Delivery Hero, which have integrated Claude Code into their development processes. Lovable, Base44, and Monday.com, three startups developing vibe-coding applications, were also present. However, outside the event, concerns are emerging. Developers are expressing themselves on online forums, criticizing the imposition of these tools by managers and highlighting the challenges posed by the additional code to review. "The only people I've heard say that the generated code is correct are those who don't read it," posted a user named pron on Hacker News.

Some developers claim that their coding skills have diminished as they delegate more tasks to AI. And researchers have warned that AI tools can produce insecure code that will make software more vulnerable to attacks.

I sat down with Katelyn Lesse, head of Claude engineering, and Angela Jiang, product lead for Claude, and asked them what they thought about concerns that a sudden influx of generated (and shipped) code without proper human oversight could lead to serious security and maintenance issues in the future. "All the old principles of good software development practices still apply. They have been applied all this time," said Lesse. "I think there are a lot of people and teams who may have lost sight of these principles right now."

And yet, as Anthropic and others push for greater automation and tools like Claude Code improve, the temptation to outsource more and more tasks, including oversight, increases. Lesse told me that some of the technical leads at Anthropic are overwhelmed by tracking all the code their teams are now producing. "Part of what is happening much more quickly is simply managing your time," she said.

"I think right now, Claude is probably as good as a mid-level engineer at writing code," she added. It still takes expert engineers to design a system and solve more complex problems, she said. "But over time, we want Claude to get better and better at all types of engineering."

Jiang nodded in agreement: "I think the end state we are trying to reach is for Claude to essentially build itself."