20 People, 30% Hit Rate: Inside Anthropic Labs, the Incubator That Hatched Claude Code and MCP

As Anthropic prepares for what is likely to be one of the most anticipated IPOs in tech history, the company’s growth story is usually told in terms of frontier models and compute deals. But a detailed Business Insider profile published September 7 of the company’s internal incubator — a rotating group of roughly 20 people called Anthropic Labs — makes a different case: the real engine of Anthropic’s commercial rise has been an unorthodox product team designed to kill bad ideas fast and keep starting over.

Labs, created by Anthropic cofounder Ben Mann in 2024, is the group behind Claude Code — arguably the definitive AI developer tool of the past two years — as well as the Model Context Protocol (MCP), the open standard for connecting AI agents to data sources and tools that is now cited at around 100 million monthly downloads. This year the team shipped Claude Design, Anthropic’s entry into AI-assisted visual creation. The unit also counts Instagram cofounder Mike Krieger, Anthropic’s outgoing Chief Product Officer, among its leaders.

A startup incubator “in the true sense”

Mann is blunt about how the team operates. “We do treat it as a startup incubator in the true sense of the term,” he told Business Insider, “and we expect most of the bets to fail.” The numbers back him up: Labs has about a 20% to 30% success rate with its ideas, some of which get folded into other prototypes or products, such as Anthropic’s Claude Chrome extension.

The team works in rapid cycles. Prototypes — internally called “bets” — come up for a “persevere or pivot” review roughly every two weeks. Bad ideas are tossed or merged into others, and the employees who built them simply jump to the next task. Good ideas “graduate” out of Labs into their own dedicated teams within the company, the same way some of Google’s moonshot bets, like Brain, eventually grew into larger divisions. Because projects usually graduate once they grow beyond four people, Labs has constant churn: members rotate out to other teams as new staffers rotate in.

The lineage is deliberate. Mann cites Bell Labs, the legendary industrial research organization, as a model, and he previously worked on a Labs-esque team at Google called Area 120 in 2018. Google’s “moonshot factory” X, which helped build Waymo and Google Brain, plays the same role in that company.

The frontier-information advantage

The key structural advantage, according to Mann, is proximity. Labs developers get early visibility into “the Anthropic milieu of what’s going well in research” — they hear directly when new models show emergent strength in a capability area, which gives them a head start on building tools for capabilities the rest of the world hasn’t seen yet.

Claude Code is the canonical example. Well before the tool’s development began in late 2024, Mann’s team had already heard from researchers that upcoming models showed signs of unusual strength in agentic coding. Mann assigned the problem to a then-new employee, Boris Cherny. Cherny’s prototype became popular internally, and Anthropic released it as a research preview in February 2025. The company’s model releases over the following year turned Claude Code into an uber-popular product — by recent accounts a $1 billion run-rate business within roughly six months — pouring gasoline on Anthropic’s growth.

Notably, Cherny — who now leads Claude Code as its own organization — had originally proposed building a much smaller code analysis tool. Mann rejected it as not ambitious 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.”

Growing the “action space”

Mann frames Labs’ mission as expanding AI’s “action space” — the set of things strong AI models can actually do in the world. The influence runs in both directions: Labs has pushed Anthropic researchers to improve audio models so they better understand Amharic, the Ethiopian language, and building Claude Design has driven new research interest in improving visual outputs.

The team’s ambitions extend beyond developer tools. Mann says he would like to see Labs contribute to breakthroughs in biological research, clean energy storage, and other real-world applications — achievements he argues would make the team’s work feel more aligned with Anthropic’s stated mission of helping humanity through stronger AI systems. “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.

The post-IPO tightrope

The profile lands at a delicate moment. As Anthropic goes public, its product strategy increasingly competes with its own customers’ software. Claude Design goes up against design tools from Adobe and Figma; Claude Code competes with development environments and coding assistants from much larger platforms.

Mike Gualtieri, an AI analyst at Forrester, told Business Insider that Anthropic faces a high-wire act: “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. It’s going to be a little dance.”

There is also the question of whether a fast-kill incubator culture survives the scrutiny of public markets. Mann noted that the Labs ethos has already spread — other product engineering teams at Anthropic now run their own “bets” groups. And a few years ago, he says, he had to convince the company it should release products at all, rather than ride philanthropic funding toward advanced AI research. That argument is definitively settled; the question now is whether a 20-person team with a 30% hit rate can keep generating billion-dollar products on a public-company clock.

For an industry obsessed with scale — trillion-parameter models, gigawatt data centers, nine-figure compute deals — the Labs story is a useful counterpoint. The unit that produced Anthropic’s most commercially important product runs on a two-week review cadence, graduates projects when they hit five people, and expects seven out of ten ideas to die. The lesson Mann offers to new hires is less about process than posture: assume the agents can do more than anyone currently believes, and build accordingly.