does verifiers allow for multiple agents in an environment?

[skunnavakkam]

trying to implement a multi-agent game, and wondering if there is a way to do this with training through verifiers

[Sohailm25]

To my understanding, verifiers doesn't natively support multiple agents in a single environment. It's designed around a single agent interacting with an environment

Current Architecture

• One model/agent generates responses at each turn
• The environment responds with feedback/observations
• This continues until a termination condition is met

Reference: verifiers/envs/multiturn_env.py:52-156 - The rollout() method orchestrates a single model's interaction with the environment

1. TextArenaEnv explicitly sets num_players=1: verifiers/envs/textarena_env.py:92,111 - Even though TextArena (the underlying library) supports multi-player games, the verifiers wrapper forces single-player mode

2. Single client/model per rollout: The rollout() signature in verifiers/envs/environment.py:293-309 accepts one client and one model parameter

3. State structure assumes single agent: verifiers/envs/multiturn_env.py:68-82 - The state tracks completion as a single trajectory, not per-agent trajectories

2 Workarounds though

Option 1: Environment-Simulated Opponent

Implement one agent as part of the environment logic

Option 2: Sequential Multi-Agent Rollouts (Self-Play)

Generate multiple rollouts with different model instances and aggregate results:

I can try to help code up examples of option 1 and 2, but not a bad idea to add a MultiAgentEnv base class @willccbb