Agentic Monte Carlo: Simulating Reinforcement Learning for Black-Box Agents

In three linesAgentic Monte Carlo (AMC) optimizes black-box LLM agents without parameter access. The method uses Sequential Monte Carlo to sample from the optimal policy by learning a value function to steer the agent, leaving the underlying model unchanged. Validated on AgentGym, AMC outperforms prompting baselines and GRPO.

Summary generated by Claude — human-verified