Anthropic Labs: 20 People, 2-Week Cycles, 80% Failure Rate — How Claude Code Emerged

Anthropic's internal Labs team, led by co-founder Ben Mann, operates as a startup incubator inside the AI company. Roughly 20 people work in two-week sprints to evaluate prototypes: continue, pivot, or kill. Mann states the idea success rate is only 20-30%, but failed explorations often contribute reusable components to other products such as Claude's Chrome extension.

Incubation Mechanics

Projects graduate from Labs when the team grows beyond four people. Both Claude Code and Claude Design followed this path, moving to dedicated product teams while Labs stays small and fluid. This mirrors historical models — Bell Labs, Google's Area 120 (which Mann joined in 2018), and Google X — but with a critical difference: tight coupling to frontier model research.

Research-to-Product Feedback Loop

Because model capabilities shift rapidly, Labs sits close to research. When researchers signaled that new models showed promise in agentic coding, Mann assigned the direction to new hire Boris Cherny. Cherny initially proposed a code-analysis tool; Mann pushed for a broader ambition, arguing that new employees must raise their targets since the ceiling of agent capabilities is still unknown. The resulting prototype became Claude Code, launched as a preview in February 2025. It writes, edits, and runs code, and each subsequent model upgrade further improves the product.

Two-Week Evaluation Discipline

Every two weeks Labs reviews each project: is it working? If not, the effort stops or its valuable pieces are absorbed elsewhere. Participants then rotate to new experiments. This cadence prevents Labs from being consumed by operational maintenance of mature products.

Bidirectional Learning

Product gaps discovered during prototyping feed back to research, shaping future model capabilities. Mann frames Labs' mission as expanding AI's "action space" — enabling models to do more in the real world. Coding requires code manipulation; design requires visual generation; cross-language use demands broader comprehension. Each product surfaces specific capability deficits.

Commercialization Tensions

Forrester analyst Mike Gualtieri notes Anthropic's tools increasingly compete with existing software vendors. Claude Design enters Adobe and Figma territory. The platform-versus-product tension grows: more built-in features reduce user tool-switching but threaten ecosystem partners. Anthropic's balance will affect platform attractiveness.

Future Horizon

Mann envisions Labs eventually tackling biology research and clean-energy storage, extending the same observe-prototype-validate loop. These remain long-term visions, but Claude Code proves one path works. The ongoing test: continuously identify the next high-leverage task as model capabilities and user needs co-evolve.