Back in January, Instagram co-founder Mike Krieger traded his CPO title at Anthropic for a hands-on role leading the company’s internal incubator, Labs. Alongside Ben Mann, he’s now tasked with turning Claude’s raw model capabilities into real products. Meanwhile, Ami Vora took over the product organization to scale Claude’s existing lineup.
https://www.youtube.com/watch?v=ktKS2MIiAmY
Labs is essentially Anthropic’s in-house startup team — it exists to bridge the gap between what the model can do and what users can actually touch. In a recent interview, Krieger said the expanded team has ditched the traditional two-week sprint cycle for something he calls “Bolts” — iterations measured in hours or days. Most team members have founder backgrounds, and the internal North Star is clear: find the “next Claude Code.”
That approach has already spawned several core products. Anthropic says Claude Code, MCP, Skills, Claude in Chrome, and Cowork all came out of Labs’ model-boundary experiments. Sources also note that Krieger’s earlier work on long-horizon agents fed into Claude Managed Agents.
Computer Use is a good early example of the process. Back in October 2024, Anthropic rolled out a computer use beta alongside the upgraded Claude 3.5 Sonnet — letting Claude see the screen, move the cursor, click buttons, and type. But Krieger says the feature was almost shelved because it was too slow and click accuracy was poor. It only made its way back to the product after later models showed dramatic performance improvements.
Compute power has become a hard constraint on Anthropic’s ability to ship products. On May 6, 2026, the company struck a deal with SpaceX for full capacity at the Colossus 1 data center, gaining over 300 megawatts of new capacity and more than 220,000 NVIDIA GPUs. They are also raising usage caps on Claude Code and the Claude API. For Anthropic, Labs finds the shapes — the compute deal makes them run.