In-the-Flow Agentic System Optimization for Effective Planning and Tool Use

ICLROral2026

Authors
Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Zou, Pan Lu
Affiliation
Texas A&M University - College Station
Venue
ICLR 2026
Track
Oral

TL;DR

We introduce AgentFlow, a trainable agentic system, and Flow-GRPO, an on-policy RL algorithm that optimizes the planner "in-the-flow" by broadcasting a final outcome reward to all steps, enabling effective long-horizon planning and tool use.

Opening excerpt from the authors’ abstract. source

Read the paper

Topics

optimization planning agent

← All ICLR 2026 Oral papers · Browse the whole archive