Anthropic pushes deeper automation with Claude Code

Anthropic used its London developer event to present a software workflow where Claude increasingly writes, tests, and revises code with minimal human intervention. The pitch landed with an audience already comfortable shipping code generated by Artificial Intelligence, even as concerns over review, security, and developer skill remain unresolved.

Anthropic’s London developer event showed how quickly Artificial Intelligence coding tools have moved from novelty to routine practice. During a stage poll, almost half the audience raised their hands when asked whether they had shipped a pull request in the last week that was completely written by Claude. When asked who had shipped a pull request completely written by Claude without reading the code at all, most of those hands stayed up. The response captured a new development culture in which generated code is increasingly accepted as standard output rather than a draft requiring close human inspection.

Anthropic framed that shift as the direction of travel for software development. “Most software at Anthropic is now written by Claude,” Jeremy Hadfield said. “Claude has written most of the code in Claude Code.” The company argued that the next step is not just using Artificial Intelligence to generate code, but letting it handle more of the full development loop. Boris Cherny said the default is moving from prompting Claude directly to having Claude prompt itself. The goal is for Claude to test and revise its own work repeatedly so developers do not even need to see many of the underlying errors.

A newly announced feature called dreaming is meant to support that model. Claude Code agents write notes to themselves about tasks, preserving useful information and lessons from mistakes. Later agents working on the same code can use those notes to get up to speed faster. Dreaming then reads across those notes, consolidates them, and identifies recurring patterns and common issues. Anthropic’s pitch is that this process will help Claude Code learn the specifics of a code base and become more effective over time.

The event also highlighted companies that have reorganized development work around Claude Code, including Spotify, Delivery Hero, Lovable, Base44, and Monday.com. Inside the conference, enthusiasm appeared widespread. Outside that setting, skepticism has grown louder. Some developers argue that managers are pushing Artificial Intelligence coding tools for productivity gains even when the result is more code to review and maintain. Others say their own coding skills are deteriorating as they delegate more work, while researchers have warned that generated code can introduce security risks.

Anthropic’s leaders did not dismiss those concerns, but they placed responsibility on teams to preserve discipline as automation rises. Katelyn Lesse said traditional software development best practices still apply and suggested some teams have lost sight of them. She also acknowledged that the pace of output is straining managers who must keep up with much larger volumes of code. Lesse said, “I think that right now Claude is probably as good as a midlevel engineer at writing code,” while adding that expert engineers are still needed for system design and harder troubleshooting. Angela Jiang described the company’s long-term aim even more plainly: “I think the absolute end state we’re trying to get to is Claude basically being able to build itself.”

64

Impact Score

EU delays key Artificial Intelligence Act deadlines

The European Union has provisionally agreed to revise the Artificial Intelligence Act, delaying several major compliance deadlines and narrowing some obligations. The changes give businesses more time to prepare while preserving the law’s overall high-risk framework.

Empirical Research Assistance automates scientific coding

Empirical Research Assistance, a system developed by researchers at Google and Harvard, automatically writes and refines scientific software for scorable research tasks. Tests showed it could outperform expert-built programs across problems including COVID-19 forecasting, neural modeling, and single-cell RNA sequencing analysis.

Google unveils new Artificial Intelligence models and personal agents

Google used its I/O developer conference to introduce updated Gemini models and personal Artificial Intelligence agents aimed at competing more aggressively with OpenAI and Anthropic. The push centers on stronger models, wider product integration, and a broader enterprise and developer pitch.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.