Agentic AI
AgentFlow improves AI planning through modular training inside live interactions
A new framework trains specialized AI modules collaboratively during multi-turn conversations, outperforming larger models on reasoning and tool-use tasks.
Researchers have proposed AgentFlow, a new approach to building AI agents that coordinate multiple specialized components while learning directly from real interactions—rather than treating the agent as a single monolithic model.
The Core Problem#
Current tool-augmented language models typically combine planning and tool calls into one large policy trained offline, which struggles with long task sequences and doesn't adapt well to new scenarios. Multi-module agentic systems promise better decomposition but most either skip training entirely or rely on offline preparation disconnected from actual conversation dynamics.
How AgentFlow Works#
The framework orchestrates four specialized modules through shared memory:
- Planner — decides what actions to take
- Executor — carries out tool calls
- Verifier — checks intermediate results
- Generator — produces text responses
These modules are trained continuously as the agent operates in real environments, rather than beforehand in isolation.
The Training Innovation#
To solve the credit-assignment problem in long-horizon tasks with sparse rewards, the team developed Flow-based Group Refined Policy Optimization (Flow-GRPO). This method converts multi-turn optimization into a sequence of manageable single-turn updates, taking each conversation's final outcome and propagating it backward to align every local decision with overall success. It also stabilizes learning using group-normalized advantages.
Empirical Results#
Testing across ten benchmarks, AgentFlow with a 7B-parameter backbone achieved substantial gains over existing systems: 14.9% accuracy improvement on search tasks, 14.0% on agentic challenges, 14.5% on math problems, and 4.1% on scientific tasks. Notably, it exceeded the performance of larger proprietary models including GPT-4o. Additional analysis showed stronger planning quality, more reliable tool-calling, and continued improvements as model size and reasoning steps increased.
