Player-Coach Teamwork: Multi-agent Collaboration for Improving LLM Reasoning

Large language models (LLMs) are increasingly deployed as autonomous agents in interactive environments. These settings require multi-step reasoning, adaptation to context, and long-horizon decision making. Methods that incorporate reasoning traces, along with reflection-based approaches, have improved robustness. However, single-agent designs still reinforce internal biases and often correct only after errors accumulate, limiting their effectiveness in real-time interactions. Multi-agent frameworks demonstrate the potential of role separation and collaboration, but typically require fixed protocols or additional training overhead. We propose a Player–Coach collaboration system that achieves conditional, real-time correction without additional training. The Player engages directly with the environment through a reasoning–acting loop, while the Coach is selectively invoked under high uncertainty to provide metacognitive feedback such as clarifying objectives, surfacing overlooked observations, or resolving confusions. A composite uncertainty trigger, combining normalized entropy and margin signals, ensures interventions occur only when needed. Experiments on the ALFWorld benchmark show that this architecture improves success rates and reduces trajectory lengths compared to single-agent baselines, highlighting the promise of uncertainty-aware multi-agent design for scaling LLMs in interactive environments.