Profile-Aware Maneuvering: A Dynamic Multi-Agent System to Robust Agentic Problem Solving

Zhitian Xie, Qintong Wu, Chengyue Yu, Chenyi Zhuang, Jinjie Gu

Keywords: Dynamic Multi-Agent System, Runtime, Stability, Maneuvering, System Identification

Abstract: The rapid advancement of large language models (LLMs) has empowered intelligent agents to leverage external tools for solving complex problems, yet this reliance introduces new challenges as extended contexts and noisy tool outputs undermine system reliability. We argue that building robust agents requires the rigor of control engineering, rather than relying on empirical prompt engineering. Drawing inspiration from predictive control in vessel maneuvering, we reframe agent design as a formal control systems problem. We first establish a baseline Multi-Agent System (MAS) where a Guard Agent acts as a simple reactive feedback controller, correcting a primary Execution Agent's errors after they occur. However, this reactive approach is fundamentally limited. Our core contribution, termed Profile-Aware Maneuvering, elevates this to a predictive control architecture. Through an automated offline 'System Identification' process, we generate an explicit, text-based 'performance fingerprint' modeling the Execution Agent's characteristic failure modes. Armed with this fingerprint, the Guard Agent evolves from a reactive critic into a predictive controller. It implements a feed-forward strategy to preemptively counteract errors before they derail the reasoning process. Experiments across a spectrum of benchmarks, including GAIA, HLE, and GPQA Diamond, validate our approach. The final Profile-Aware MAS demonstrates the hallmarks of a well-controlled system: it dramatically reduces performance variance while simultaneously boosting accuracy, and it minimizes the gap between its potential and single-pass performance. This superior performance and stability culminated in our system achieving a score of over 81 on the GAIA leaderboard. Our findings advocate for a paradigm shift: from the empirical art of prompt engineering to the principled science of control theory for designing predictable and trustworthy intelligent agents.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 5588