Worlds built for agents, not adapted for them.
Licensed games give an agent realism it could not get anywhere else. What they cannot give is a world whose every uncertainty was put there on purpose. So we build those too.
benned Worlds is a set of text-and-state environments built in-house for long-horizon planning, alongside the licensed commercial games benned operates for AI labs. The first is BuildSite: an agent runs a construction project against suppliers with uncertain lead times, crews of varying skill, weather and defects that only surface later. It is procedurally generated, fully reproducible per seed, and reachable through both a structured interface and plain text.
A project that punishes you nine weeks later.
A construction project is a planning problem with teeth. Materials have lead times you only half know. Crews differ in speed and in what they get wrong. Weather removes days you had counted on. And a defect built in early stays invisible until something later fails on top of it.
An agent that reacts well step by step does poorly here. Finishing on time and on budget takes a plan that survives contact with all four at once.
- Procedurally generated and fully reproducible per seed
- Structured and plain-text interfaces, for RL and language-model agents alike
- About 85 complete episodes per second on one CPU core
- A rule-based baseline finishes most projects but rarely on budget, so there is clear headroom to measure against
- Scenarios, difficulty and scoring designed with the lab
Realism and control are different instruments.
A licensed simulator gives an agent a world that behaves like the real one, with all the mess that implies. A purpose-built world gives a lab a knob on every source of that mess. Research usually needs both, which is why we run both rather than choosing.
- Same isolationDedicated hardware per lab, no agent data stored, keys held by the lab.
- No licence to negotiateThe world is ours, so access starts without a studio agreement in between.
- Cheap to runText-and-state rather than rendered, so throughput is a CPU question, not a GPU budget.
About benned Worlds
What is a long-horizon planning environment?
One where the consequence of a decision arrives long after the decision — ordering the wrong material in week one shows up as a delay in week nine. It tests whether an agent can plan under uncertainty and hold a plan together, rather than react well step by step.
How is BuildSite different from a licensed game?
A game was built to be fun for a human and then made agent-ready. BuildSite was built to be an environment: it is deterministic per seed, runs far faster than real time, and every source of uncertainty in it is deliberate and tunable.
Can a lab shape the tasks in benned Worlds?
Yes. The world generator and the scoring are ours to change, so scenarios, difficulty and what counts as success can be designed with the lab rather than worked around.