Real-world work,simulated.
Simulated workplaces and worlds with fixed rules, a seeded start and a score computed from the end state — from a 24-tick shift to hundreds of turns. Use ours, or have us build yours.
Loading dock
OperationsSchedules trucks, doors and crews through a shift where delays cost money. Work that looks like a real job.
Procedural world
Long horizonA new seeded map every day. Explore, fight and manage resources over hundreds of turns, with hidden state the agent has to discover.
Embodied 3D
Perception · controlResearch harnesses that put models in control of 3D game worlds — missions, recovery, navigation — recorded on video with every model call logged.
More operations sims
Dispatch · logistics · planningFurther operations environments built on the same kit — one folder, one scorer contract, the same replay and capture.
Every environment shipsthe same five parts.
That contract is what makes results comparable across models, versions and teams — and why a custom world takes days, not months.
- 01RulesA deterministic engine. Server-held state; agents only see what they've discovered.
- 02Scenario bankFixed, versioned seeds for fair comparisons, plus held-out seeds that never leave the building.
- 03Baselines0 = a no-plan policy, 100 = the hindsight optimum. A scripted expert sits in between.
- 04ScorerCode computes the score from the end state. No rubric, no judge to persuade.
- 05ReplaySeed + action log reproduces the run and its score exactly — the basis of every claim.
Need an environment thatlooks like your task?
Tell us the job, the agents involved and where they fall short. We'll tell you whether an environment can answer it.
Contact us →