What the platform provides
A complete experimentation spine—from pluggable worlds and model agents to live telemetry, evaluation, replay, and environment generation.
Environment, agent, evaluator, and runtime plugins
Typed registries keep concrete worlds, policies, scoring logic, and execution backends replaceable. The core coordinates them without importing game rules, chemistry logic, or renderer code.
Default agents and real model calls
Every showcase has a reliable local agent for instant demos. The server-side OpenRouter plugin adds curated selectable models and OpenRouter Auto through one normalized agent contract, with different models assignable to competing participants.
Observable runs
Live status, active participant, actions, validation, tool use, latency, token usage, cost, scores, errors, and environment state are exposed through normalized events and rendered in each world's purpose-built interface.
Guardrails and reproducibility
Runs can enforce step, token, cost, retry, and time limits. Environment configuration, model identity, seeds, accepted actions, evaluator output, snapshots, and terminal state travel with the persisted record.
Human-in-the-loop operation
Interactive worlds accept human actions through the same external-action endpoint used by the orchestrator. Human control never bypasses environment schemas, rules, event emission, evaluation, or persistence.
Codex environment workshop
The Build workbench turns a creative brief into an isolated candidate plugin, validates package structure and ArenaOS lifecycle contracts, previews artifacts, and requires approval before registration. Production environments are not mutated by failed builds.