Skip to main content
AWP is a v0.1 draft. Propose changes through the RFD process.
The Agent World Protocol (AWP) standardizes the boundary between a mind and a world. A mind is any policy that maps observations to actions — a multi-modal LLM, a diffusion policy, a classical planner. A world is anything that produces observations and accepts actions — a game engine, a robot body, a browser, physical reality reached through sensors and actuators. MCP standardizes agent↔tools. A2A-style protocols standardize agent↔agent. Gym standardized the training loop, and ROS standardized robot middleware. AWP fills the missing seam: a protocol for embodied interaction at inference time, spoken between an agent and a world that share an environment, where worlds are self-describing, time is real, actions have duration, and safety is enforced at the boundary.

The five primitives

Where to start

Read the specification

The normative v0.1 draft: transports, sessions, the loop, and safety.

Why AWP?

The gap it fills next to MCP, A2A, Gym, and ROS.