Environments to act in.
Live tools, real codebases, and stateful workflows. A place for agents to try, fail, and try again.
Environments for intelligence
Intelligence
learns by
doing.
RL environments and high-quality training data at scale for AI labs building models that can do real work.
Learning through experience
Our approach
Real work is
the curriculum.
Useful intelligence takes more than the right answer. It takes the ability to act, adapt, and follow through.
We turn complex work into training experiences: environments with persistent state, tasks that unfold over time, and rewards grounded in what actually happened.
Data to learn from.
Tasks and trajectories across coding, science, AI research, and professional work, built around the capabilities your model needs.
Outcomes to verify.
Programmatic checks, deterministic tests, and task-specific rubrics that make progress measurable.
Domains
The world is
more than a prompt.
From long-horizon coding to scientific discovery and AI research, we build environments for work that takes more than one step.
A codebase that keeps evolving.
Agents edit, test, and debug across successive requirements. Each change becomes part of the next challenge.
- Real repositories
- Docker environments
- Programmatic tests
Illustrative task
Modernize a Python package without breaking its public interface.
Workspace
src/tests/pyproject.toml- Inspect the codebase
- Implement the migration
- Run tests & debug
- Verify compatibility
Behavior preserved · Build reproducible
Research & open work
Make the next
capability measurable.
BenchmarkPre-release
Multi-step-SWE
One codebase. Successive requirements. A benchmark for coding agents working on software that keeps evolving.
Paper & leaderboard coming soonExplore the benchmarkKlavis AI
Building the conditions
for better intelligence.
We work at the intersection of agent infrastructure, data, and evaluation. Our focus is practical: what does a model need to experience to become more capable?
Y Backed by Y Combinator