paper

About

AgentFlow came out of Stanford to attack credit assignment in long-horizon agent runs, where a frozen LLM loops on failed tool calls and never learns from the outcome. The system splits work across a planner, an executor, a verifier, and a generator coordinated through shared memory, and trains only the planner by converting each multi-turn trajectory into single-turn policy updates with group-normalized advantages. After training, an agent that once repeated the same failed tool call instead self-corrects and changes strategy. The framework targets researchers studying trainable agentic systems, with code, project page, and a demo space available.