High signal environments
for agents
Curated by human experts from diverse, high-value professional domains. All tasks are verifiable, expert verified, using real data — things humans would get paid to do.
Our Work
What we've built
SkillsBench
The first evaluation framework measuring how skills (custom instructions) work for AI agents. 84 expert-curated tasks across diverse, high-GDP-value domains. The first dataset measuring how powerful models are at using skills.
PokemonGym
First open-source harness for any LLM and agent to play Pokemon Red and Blue. Tests vision, reasoning, planning, memory, and sequential decision-making. Featured in the Gemini model launch.
BenchFlow Hub & Runtime
The first protocol for agent and benchmark unification. HuggingFace for benchmarks and RL environments. One-line setup for 60+ benchmarks spanning NLP, web agents, code, medical AI, and more.
Backed by
Jeff Dean
Arash Ferdowsi
Dropbox
Eugene Yan
Amazon
Founders, Inc.
A16z
Scout Fund