Environment contract
A world implements a typed lifecycle, owns authoritative state, and declares the actions it accepts.
Lifecycle
Initialize once, reset per episode, process validated actions through step, expose state, and close cleanly.
initialize(context)
reset(input)
step(action)
getState()
close()Schemas and capabilities
Action and observation schemas make integrations inspectable and allow the core to reject malformed actions before they reach world logic. Metadata declares deterministic, real-time, multi-agent, rendering, snapshot, and seeding capabilities.
Authority
Rules, legal actions, rewards, terminal conditions, and canonical state belong in the environment plugin. A web renderer may animate the state but must never compute a competing outcome.
Participants
Multi-agent worlds identify the active participant in observations. The orchestrator routes the next action to that participant's registered agent or waits for a normalized human action.
Replay
Renderable environments expose serializable state and stable semantic actions so ArenaOS can persist frames and reconstruct behavior without rerunning agents or external models.