Realistic game worlds, licensed from the studio, private to your lab.
You license directly from the studio that made the game. benned turns it into an environment your agents can use from day one and runs it on hardware that serves only you.
A licensed commercial game, made agent-ready and run as a private environment for one lab. Pixels or structured game state in; keyboard, mouse or discrete actions out, through a Gymnasium-style API. Fixed seeds, snapshots and instant resets, faster than real time where the engine allows, with scored tasks designed with the lab. It runs on dedicated hardware no other customer touches, and benned stores no observations, actions or trajectories — only instance hours and concurrency, for metering.
Works out of the box
A Gymnasium-style API. Pixels or game state in; keyboard, mouse or structured actions out. No game-side code for your team.
Tasks and scoring
Scenarios with automatic evaluation, designed with you, so you can track learning per run and per checkpoint.
Fast and repeatable
Faster than real time where the engine allows, fixed seeds, snapshots and instant resets. Thousands of concurrent instances.
Customisable worlds
Maps, rules, weather, difficulty and starting states varied per seed, so agents never see the same world twice.
# pip install benned import benned env = benned.make( "haulage-sim/deliver-load", obs="pixels", seed=42, speed=8.0) obs, info = env.reset() for step in range(10_000): obs, reward, done, trunc, info = env.step(agent.act(obs)) if done or trunc: obs, info = env.reset() # runs only on your lab's dedicated hardware
Priced per world, per lab, per agent.
Every engagement is quoted per game. The structure is always the same.
Integration
Making the game agent-ready for your lab, including custom tasks and scoring.
Managed service
Dedicated hardware, uptime, updates and support.
Usage
Scales with how many instances you run at the same time.
benned Worlds: environments built for agents
Alongside licensed games we build our own text-and-state worlds for long-horizon planning. The first, BuildSite, puts an agent in charge of a construction project: suppliers with uncertain lead times, crews of varying skill, weather and hidden defects.
- Procedurally generated, fully reproducible per seed
- Structured and plain-text interfaces for RL and language-model agents
- About 85 complete episodes per second on one CPU core
- Rule-based baseline finishes most projects but rarely on budget, leaving clear headroom
Before you get in touch
Can benned see or keep our agents' data?
No. Trajectories, observations and actions stay in your environment. benned meters usage only. Each lab runs on physically separate hardware.
Who do we contract with?
You license the game from the studio. benned operates the environment as the studio's technical partner under a separate service agreement.
Can it run inside our own infrastructure?
Yes, on request. The default is dedicated hardware in a Tier III facility in the Netherlands.
Which observation and action spaces are supported?
Pixels with keyboard and mouse for every game, and structured game state with discrete actions where the engine integration allows it.
Start with one game and one pilot.
Studios: tell us which titles you could license. Labs: tell us what your agents need to learn. We reply within two working days, and nothing you send goes further than us.