> ## Documentation Index
> Fetch the complete documentation index at: https://www.agentworldprotocol.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Multi-agent worlds

> Many sessions in one world, and where A2A takes over.

Each agent opens its own session with independent grants and envelopes (AWP-MA-001). Agents perceive each other as entities tagged `controlled_by: "session"` (AWP-MA-002).

Design choices that work well:

* **Communication as world actions.** A `say` action and `heard` events keep inter-agent talk inside the audit log and the world's physics (range, interception) — usually better than out-of-band channels for games and sims.
* **A2A out of band** when agents must coordinate privately or across worlds; AWP deliberately does not define this ([multi-agent](/spec/multi-agent)).
* **Spectator sessions** (AWP-EMB-004) for referees, evaluators, and dashboards.
* **Shared control** with declared arbitration for human-takeover setups (AWP-MA-003) — see [Human in the loop](/guides/human-in-the-loop).
* **Shared lockstep time.** Declare `tick_authority: "barrier"`: the world advances once every bound lockstep session has called `world.tick`, and every session receives the same advance (AWP-TIM-012). A stalled agent then blocks the round — give barrier worlds a per-tick timeout that treats a silent session as having ticked, and document it in the manifest.
* **Administrative mutations.** `world.reset`, `world.restore`, and (where the world chooses) `world.tick` require the `admin` grant class (AWP-PRM-005). Grant them to the harness session, not to competing agents; a reset cancels every session's in-flight actions (AWP-PRM-006).