Anthropic Labs

Sources#

Summary#

Anthropic's internal incubator — "a lab inside a frontier lab." Dan Carey, a PM within Labs, describes it as a bet factory: at any time ~a dozen small teams explore the frontier of what the models can do, make bets on whether something will work, run experiments to find out, double down on what works, and fold what doesn't. The operating model is lean-startup methodology run at unusual speed ("the biggest difference is the speed at which we run our loops"). Labs is the origin point for several products that escaped into general use: Claude Code, MCP, Skills, Claude Design, the Claude-in-Chrome extension, and hands-free audio work.

The bet-factory operating model#

Lots of exploration, few high-conviction releases. Many small bets; a small number become "moonshots" that change how people work with the models. "Most labs bets do not make it past [early exploration]. Most things we end up folding."
Start as one person. Almost every bet begins as a single person "with their good buddy Claude" exploring a hard idea, looking not for a finished product but for "that little hint of heat" — a sparkly glimmer worth building on. Early exploration is usually hours to a few days.
Scale tiny, deliberately. A promising spark scales "all the way to three people" — kept small to minimize collaboration overhead — and only "all the way up to five" ahead of launch (see Engineer PM Convergence, Managers as ICs).
Ship daily; don't predict the future. "We do not try to predict the future… we ship, we watch, we learn" — run the loop repeatedly rather than forecasting a ten-year roadmap (see Compounding Loop Optimization).
Spark over plan. Bets start from a prototype + a Slack post, not a PRD, vision doc, or OKR meeting (see Prototype Over PRD).

Rituals and practices#

Pitch-offs — periodic brainstorm sessions where people try to "nerd-snipe" colleagues into joining their bet. The first Claude Design pitch-off (where 100% of second-half pitches were made live with the tool) was the proof point that convinced the team to take it to market.
Talk to users and researchers daily — Carey's two favorite prompts: to users, "please complain at me"; to researchers, "what have you been surprised by lately?" Both are sources of new bets.
Slack as the operating surface — bets are shared, critiqued, and roadmapped in Slack.

History (reconciling the two rounds)#

Per Anthropic's entity page and Boris Cherny: a first incarnation of the Labs incubator formed in late 2024, produced Claude Code, MCP, and the desktop app, and "disbanded after launches." A round 2 is led by Mike Krieger (former Instagram co-founder, who runs Anthropic's product side). Carey's account — Labs as an ongoing bet factory shipping Claude Design, Skills, and more in 2026 — describes this current incarnation. Carey does not assign each product to a specific round; the through-line is the bet-factory model itself.

Connections#

Anthropic — parent company; Labs is its internal incubator
Dan Carey — PM within Labs; primary source for this page
Claude Design — current flagship Labs bet (built by ~3 people in ~10 weeks)
Claude Code — earlier Labs product; the velocity it created is why Claude Design was started
Cowork — sibling Anthropic agent product
MCP and Computer Use — MCP originated in Labs
AI Native Product Cadence — the bet-factory cadence is the same ship-fast/research-preview pattern Cat Wu describes at team level
Engineer PM Convergence — the 1→3→5 "everyone does everything" team shape Labs runs on
Compounding Loop Optimization — the per-bet operating discipline
Prototype Over PRD — Labs bets start from prototypes, not planning docs
Build for the Next Model — Labs prototypes "the thing that almost works" and lets the next model close the gap

Sources#

Designing with Claude: From prompt to production — Dan Carey, "Designing with Claude," May 2026 (bet-factory model, rituals, team scaling)
Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next — Labs round 1 history; Mike Krieger leads round 2