Multi-agent (MARL)¶

Multi-agent support is new and has only had superficial testing compared to the single-agent API.

make_parallel_env() provides the supported PettingZoo ParallelEnv facade.

from mudgym import make_parallel_env

env = make_parallel_env(agents=4)
obs, infos = env.reset(seed=123)
obs, rewards, terminates, truncates, infos = env.step({agent: "look" for agent in env.agents})
env.close()

possible_agents is the fixed list player_0 through player_3. Each ID represents one player life during the current environment episode. A finished ID is removed from agents and is never revived within that episode. Its final observation and raw response are returned on the terminating or truncating step. A full reset() replaces the world and starts a new lifetime context for all IDs. Observation and action spaces retain their object identity across lives and resets.

This follows the Parallel API's fixed possible_agents contract and its official no-revival test. MudGym does not generate an unbounded sequence of PettingZoo agent IDs.