![[2602.13671] MAS-on-the-Fly: Dynamic Adaptation of LLM-based Multi-Agent Systems at Test Time](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.13671] MAS-on-the-Fly: Dynamic Adaptation of LLM-based Multi-Agent Systems at Test Time
Summary
The paper presents MASFly, a novel framework for dynamic adaptation of LLM-based multi-agent systems at test time, enhancing task performance and adaptability.
Why It Matters
This research addresses the limitations of static multi-agent systems by introducing a dynamic adaptation mechanism that mimics biological adaptability. This advancement is crucial for improving the effectiveness of AI in complex, real-world tasks, where flexibility and responsiveness are essential.
Key Takeaways
- MASFly enables dynamic adaptation of multi-agent systems during deployment.
- The framework utilizes a retrieval-augmented SOP instantiation mechanism for customization.
- A Watcher agent provides real-time monitoring and interventions based on past experiences.
- MASFly achieved a 61.7% success rate on the TravelPlanner benchmark, showcasing its effectiveness.
- The approach enhances robustness and adaptability in complex task environments.
Computer Science > Multiagent Systems arXiv:2602.13671 (cs) [Submitted on 14 Feb 2026] Title:MAS-on-the-Fly: Dynamic Adaptation of LLM-based Multi-Agent Systems at Test Time Authors:Guangyi Liu, Haojun Lin, Huan Zeng, Heng Wang, Quanming Yao View a PDF of the paper titled MAS-on-the-Fly: Dynamic Adaptation of LLM-based Multi-Agent Systems at Test Time, by Guangyi Liu and 4 other authors View PDF HTML (experimental) Abstract:Large Language Model (LLM)-based multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. However, existing works often rely on manual designs or "one-size-fits-all" automation, lacking dynamic adaptability after deployment. Inspired by how biological systems adapt, we introduce MASFly, a novel multi-agent framework enabling dynamic adaptation at test time. To adapt system generation, MASFly employs a retrieval-augmented SOP instantiation mechanism that leverages a self-constructed repository of successful collaboration patterns, enabling the LLM to assemble customized MASs for new queries. For adaptive execution, MASFly incorporates an experience-guided supervision mechanism, where a dedicated Watcher agent monitors system behaviors with reference to a personalized experience pool and provides real-time interventions. Extensive experiments demonstrate that MASFly achieves state-of-the-art performance, most notably a 61.7% success rate on the TravelPlanner benchmark, while exhibiting strong task adaptability and robustne...
