Agents3h ago
COSPLAY Framework Improves Long-Horizon Agent Reasoning
The COSPLAY framework introduces a co-evolution system where an LLM decision agent retrieves actions from a learnable skill bank. This architecture solves the problem of inconsistent decision-making in complex, multi-step game environments. By retaining and reusing structured skills across episodes, the system overcomes delayed rewards. Practitioners can now build agents that maintain stability over longer horizons.