Inside Anthropic's London Pitch for Coding's AI-First Future

The News

Nearly half the room at Anthropic's Code with Claude conference in London raised their hands when asked who had shipped pull requests written entirely by Claude. Most kept them up for the follow-up: who had merged the code without reading a line of it.

The two-day event on May 19-20, running in parallel with Google's I/O in Palo Alto, drew engineers from Spotify, Delivery Hero, Lovable, Base44 and Monday.com. Claude Code lead Boris Cherny, engineers Jeremy Hadfield and Ravi Trivedi, engineering chief Katelyn Lesse and product lead Angela Jiang pitched an autonomous coding agent that has matured into a mainstream developer tool, one that Anthropic is itself running on.

"Most software at Anthropic is now written by Claude," company leadership told the audience. Hadfield went further: "Claude has written most of the code in Claude Code." A new "dreaming" feature, unveiled two weeks earlier, lets agents take self-notes, consolidate context across long sessions and surface recurring patterns. Claude 4.6 and 4.7, released in February and April, supply the horsepower.

Why It Matters

The London show was the loudest articulation yet of a stance Anthropic has been edging towards: an agent should be allowed to cook without a human watching the pan. Trivedi's three-word slogan, "let it cook", captured a philosophy that even GitHub Copilot did not dare advance when it launched in October 2021. Copilot was framed as autocomplete. This is framed as a coworker.

Lesse's claim that Claude is "probably as good as a midlevel engineer" is a different category of assertion to the narrow-benchmark productivity gains Microsoft cited for Copilot in 2023. Jiang's end state, in which Claude is "basically building itself", points to a labour structure that did not exist three years ago.

The flip side: hands raised in a ballroom are also hands lifted off the wheel. Code merged unread is code whose security flaws, licence violations and logic bugs do not surface until production, a risk that falls hardest on industries which buy software wholesale.

Indian Angle

India's software services industry is the world's largest such buyer and seller of code, making it the constituency for whom Anthropic's London message lands hardest. TCS, Infosys, Wipro, HCLTech and Tech Mahindra together employ well over two million engineers, the bulk in the entry- to mid-level grades that Lesse described as Claude's parity zone. Infosys' Topaz and TCS' generative AI suite have positioned around this shift, but Anthropic's claim that most of its own code is now written by the model is the executive-deck slide that will accelerate every procurement conversation.

For Indian product startups the calculus is brighter. Razorpay, Cred, Postman, Zerodha, Groww and Meesho have been candid about heavy Claude usage, and a let-it-cook workflow lowers shipping costs for teams that have always done more with smaller headcounts. India's pool of roughly 5.4 million developers, per GitHub's Octoverse, is on course to overtake the United States within the next year.

The harder question lands on Sarvam AI and Ola's Krutrim. A homegrown coding-grade model now needs agentic execution and dreaming-style memory, not just multilingual reading comprehension. Neither has publicly demonstrated that today.

FAQ

How does Claude's "dreaming" feature work?

Between tasks, Claude agents write notes to themselves, consolidate information from prior sessions and identify recurring patterns. The aim is working memory for long-running coding agents, so a refactor spanning days does not lose its earlier context. Anthropic unveiled it about two weeks before the London conference.

Which Indian companies use Claude in engineering?

Razorpay, Cred, Zerodha, Groww, Postman and Meesho have spoken publicly about deploying Claude or Copilot inside their engineering teams. Indian IT services majors Infosys, TCS and Wipro have built proprietary AI-coding overlays on top of multiple foundation models, including Claude, marketed to enterprise clients.

Should Indian IT services investors worry about headcount?

Nasscom data shows entry-level IT hiring slowed materially through 2024 and 2025. If Claude is at midlevel-engineer parity, the pyramid base of the Indian services model, the juniors writing routine code, is the layer most exposed to substitution rather than augmentation.

Where can I read the original report?

MIT Technology Review's Will Douglas Heaven attended Code with Claude in London and filed the feature dated May 21, 2026. It expands on the attendee show-of-hands, the dreaming announcement and the broader argument about agentic coding's risks and adoption curve.

This story was reported by MIT Technology Review, surfacing in The Download and expanded in the underlying Code with Claude feature.