AgentFlow

About

AgentFlow approaches agentic systems as trainable rather than hand-prompted: four modules (Planner, Executor, Verifier, Generator) share an evolving memory and coordinate through tool calls — python execution, web and Wikipedia search — while the Flow-GRPO algorithm optimizes the planner online against sparse long-horizon rewards. The published results show a 7B backbone outperforming GPT-4o on 10 benchmarks spanning search, agentic, math, and science tasks, with gains of roughly 14-15% on the agentic suites. Models come from OpenAI, Google, DashScope, DeepSeek, Together, or local vLLM, and training configuration lives in train/config.yaml. Researchers in agentic RL and tool use are the users, and the paper was accepted at ICLR 2026.