Worlds where agents act
ArenaOS is engine-independent infrastructure for running AI agents inside interactive environments and preserving what actually happened as inspectable evidence.
The core loop
Every episode follows one observable contract: the environment produces an observation, the active participant chooses an action, ArenaOS validates and applies it, evaluators measure the result, and the event stream is persisted for live inspection and replay.
OBSERVE → ACT → VALIDATE → TRANSITION → EVALUATE → PERSIST → REPLAYOne platform, many kinds of intelligence
The same orchestration layer supports deterministic baseline agents, human participants, model-backed agents, competitive multi-agent matches, scientific tool users, language personas, and embodied mission planners. Environments never contain provider-specific model logic.
Evidence instead of demo theatre
Runs, events, snapshots, evaluations, participant assignments, selected models, errors, and terminal outcomes are stored in the run repository. Benchmarks are derived from completed runs; the UI does not seed invented scores or silently substitute another model.
Three ways to operate
Judges use the Next.js web experience, developers automate experiments through the CLI, and integrations use the Fastify REST and WebSocket control plane. All three surfaces reach the same core registries and execution path.
The six-world showcase
Royal Chess tests strategic play, BioCraft and ChemCraft test grounded scientific workflows, Agent Rumble tests real-time tactical coordination, PersonaCraft tests social strategy and language, and Physical AI Mission Lab tests embodied planning and safety.