Skip to main content
Modern agent stacks have standardized almost every interface except the one where an agent meets a world. Each of these has asynchrony, introspection, or security of its own. None of them standardizes the combination embodied interaction needs: observation freshness as a first-class quantity, actions with duration and preemption over a shared physical state, per-session authority over an embodiment, quantitative limits enforced at the boundary, and a portable session record. The consequences show up everywhere:
  • Every simulator, game, and robot stack invents a bespoke agent integration.
  • Agent code trained in simulation must be rewritten to run on hardware.
  • Safety is enforced (or not) ad hoc inside each integration.
  • Sessions are not portable, replayable, or auditable in a common format.

What AWP changes

Self-describing worlds. A world advertises its observation channels, action schemas, time models, and safety policy in a manifest. A generic agent can enter a world it has never seen and discover: here I have two arms, a camera at 30 fps, and these 14 action types. One agent, sim and real. The protocol defines two time models — lockstep for pausable worlds, streaming for reality. The schemas stay identical, so the same agent binary can train against a simulator adapter and deploy against a robot adapter (the planned awp-sim and ROS 2 bridge are the reference designs). See Sim-to-real. Safety at the boundary. Scoped grants, server-side envelopes, approval hooks, and a normative audit log live in the protocol itself, not in each integration. See the safety model. Sessions as data. Every session is a timestamped log of observations and actions — replayable for debugging, reusable as evaluation benchmarks and training data. See Record and replay. AWP does not replace any of the standards above — it composes with them. An AWP world can expose MCP tools; an AWP robotics adapter sits on ROS topics; agents inside an AWP world can talk A2A to each other. See the Compare section.