Skip to main content
No. AWP defines the wire boundary between any agent and any world. Frameworks (and raw LLM loops) sit on the agent side and speak the protocol.
No — see Compare. Tools stay on MCP; intra-robot plumbing stays on ROS; AWP is the seam between a mind and an environment.
Deliberately out of core. Worlds MAY publish reward on an ordinary channel; the protocol stays reward-agnostic so the same world serves RL training, evaluation, and deployment.
They are siblings in genre, different in boundary: ACP connects code editors to coding agents; AWP connects agents to worlds. In ACP both parties are software minds; in AWP one party is an environment enforcing physics and safety. See AWP vs ACP.
Out of scope by design. Thousands of parallel env instances inside one trainer process are an in-process, training-time pattern; train vectorized in-process, then evaluate and deploy over AWP. A trainer MAY expose a single instance over AWP for inspection.
Yes — the game profile covers avatars, spectators, multi-session play, and match replays.
Not yet.
Not yet — the suite isn’t published.