Claude Code began as a bet in Anthropic’s Labs incubator

Anthropic’s Labs team, a rotating group of about 20 employees led by cofounder Ben Mann, expects most of what it builds to fail. Mann puts its hit rate at 20% to 30%, he told Business Insider in an interview published September 7. The hits include Claude Code, the Model Context Protocol and this year’s Claude Design.

“We do treat it as a startup incubator in the true sense of the term,” Mann said, “and we expect most of the bets to fail.” Some of the misses don’t disappear entirely. They get folded into other prototypes or products, and the Claude Chrome extension is one example. “We’ve had more hits than I would have expected, and the hits that we’ve had have gone really far,” he added.

Persevere or pivot

Labs works in cycles. Roughly every two weeks, each prototype, which the team calls a “bet”, comes up for a “persevere or pivot” review. Bad ideas are dropped or merged into other ideas, and the people on them move to another task. Good ideas “graduate” into their own teams elsewhere in the company, much as Google Brain grew out of a Google moonshot into a larger division. A project usually graduates once it needs more than four people. That means Labs is always losing members to the teams it spins out and pulling in new staff to replace them.

The team began in mid-2024 with two members, The Verge reported. In January 2026 Anthropic decided to expand it, with a goal of doubling Labs headcount within six months. As part of the same reshuffle, Instagram cofounder Mike Krieger stepped down after two years as chief product officer to co-lead Labs with Mann, and head of product Ami Vora took over his duties. “The speed of advancement in AI demands a different approach to how we build, how we organize, and where we focus,” president Daniela Amodei said at the time. “Labs gives us room to break the mold and explore.”

In its own announcement, Anthropic described the process in three motions: tinker at the edge of what Claude can do, test unpolished versions with early users, then scale what lands into products customers can rely on.

Mann describes the team’s main advantage as proximity. Labs developers have access to the “Anthropic milieu of what’s going well in research,” so they hear early how models are improving and can start building before a capability ships.

How Claude Code got started

Researchers had told Mann’s team that new models showed signs of strength in agentic coding, where the model writes, edits and runs code on its own across many steps, well before work on the tool began in late 2024. Mann gave the task to Boris Cherny, then a new hire. Cherny’s first proposal was a code analysis tool, and Mann told him the idea wasn’t big enough.

“Especially when people are new to the company, they need to change their level of ambition to much higher,” Mann said. “Because a lot of what the agents can do, nobody knows.”

Cherny’s prototype caught on internally, and Anthropic released it as a preview in February 2025. Model releases over the following year drove Claude Code’s growth, and Anthropic says it went from research preview to a billion-dollar product in six months. Cherny now leads Claude Code. MCP, the other early Labs project, has reached 100 million monthly downloads, according to the company. The cycle has also gotten faster: The Decoder reported that Cowork, which brings Claude Code’s capabilities to office work, was built in Labs in a week and a half.

Widening the action space

Mann places Claude Code inside a broader Labs goal: growing AI’s “action space”, the range of things strong models can actually do in the world. That sometimes runs in reverse, with product work setting research priorities. Labs has pushed researchers on other teams to improve audio models so they better understand Amharic, the Ethiopian language, and building Claude Design has generated research interest in better visual outputs.

The approach has spread. Other product engineering teams at Anthropic now run their own “bets” groups, Mann said.

Getting here took some convincing. A few years ago, Mann said, he had to persuade Anthropic to release products at all rather than ride philanthropic funding toward advanced AI. He created Labs in 2024 and cites Bell Labs as a model. In 2018 he worked at Area 120, a similar incubator inside Google. Google’s moonshot factory, now called X, helped build Waymo and Google Brain.

“A little dance” after the IPO

Labs is likely to matter more as Anthropic moves toward a public listing. Mike Gualtieri, an AI analyst at Forrester, said the company faces a high-wire act as it ships tools that compete directly with software its business customers already use. Claude Design, for example, competes with Adobe and Figma.

“As they build the platform, they’re going to likely create features that make it more attractive to use their platform than someone else’s,” Gualtieri said. “It’s going to be a little dance.”

Mann has his sights set elsewhere. He would like Labs to contribute to breakthroughs in biological research and clean energy storage, work he says would tie the team more closely to the company’s mission. “The ability to accelerate the hard sciences is extremely exciting and very, very important for sort of starting to deliver more benefits to people,” he said.