GraphPlanner: Multi-Agent Routing Generation
新智元 · wechat · 2026-07-12
Researchers at UIUC propose **GraphPlanner**, evolving traditional LLM Routers from "which model to choose" to "what kind of multi-agent workflow to generate." ### Core Concept - Each decision step involves not just selecting an **LLM backbone**, but also an **Agent Role**. - Roles default to three categories: - **Planner**: Decomposes complex problems - **Executor**: Answers the original question or sub-problems - **Summarizer**: Aggregates intermediate results - The routing process is modeled as a sequential decision problem, dynamically determining whether to decompose tasks, how many models to invoke, and what role each model plays based on task complexity. ### Memory Mechanism GraphPlanner introduces a heterogeneous graph memory network, **GARNet**, structuring historical interactions into a graph: - query - response - LLM/role combination - accuracy / cost This links with the current workflow state, leveraging past experience to guide the next routing step. ### Training and Optimization - Workflow generation is modeled as an **MDP**. - Trained using **PPO**. - Rewards factor in both task accuracy and invocation cost, managing the accuracy-cost trade-off. ### Experimental Results - Evaluated across **14 tasks and 6 domains**. - Outperforms baselines in both fixed and freely generated workflows. - In the second phase, it achieves an average accuracy improvement of roughly **9.3%** over the strongest baseline. - Incurs lower GPU computational overhead during training. - Demonstrates stronger generalization to **unseen tasks** and **unseen models**. ### Conclusion GraphPlanner's value lies in transforming the Router from a mere "model selector" into an **agentic workflow planner** capable of dynamically constructing collaborative processes based on the task.
More from coding & agent
- AI makes software easier to build, but it also lowers the floor on quality — paw_lean · 2026-07-21
- Fable coding run costs $6.69 for 67 lines of code in a 4-minute job — bytebot · 2026-07-21
- Cursor writes better code, but ChatGPT can still control the computer — vista8 · 2026-07-21
- Agents can remember facts, but still forget how to do the job — No_Advertising2536 · 2026-07-21
- Agent skills for project downgrade and troubleshooting tested in CLAD on LS 5.22 — stspanho · 2026-07-21
- Open-source B-roll skill turns scripts into 5-second vertical clips with Codex and Gemini — yangyi · 2026-07-21