Multi-agent Imitates Enterprise Dynamics

ICLR 2026 Conference Submission12807 Authors

Keywords: multi-agent systems;enterprise dynamics;behaviour simulation;large language models;

Abstract: Imitating enterprise dynamics characterized by volatility, long-horizon coordination, and decision-making offers executives and operational teams a structural understanding of the organization at a much lower cost. Existing LLM-based agent systems have great potential to simulate these activities, yet they still face three challenges to understand the dynamics at the enterprise scale in terms of structure, strategy, and operation. This motivates us to propose TaskWeave, a novel LLM-based multi-agent framework that aims to imitate complex enterprise dynamics. Inspired by theories in the fields of control and business management, our TaskWeave operates at three levels: strategic (executives, e.g. generating phase plans), tactical (coordination, e.g. scheduling resource allocation), and operational (task execution, e.g. leveraging multiple tools), simulating modern enterprise dynamics in an end-to-end manner. Our TaskWeave instantiates an IT company to simulate year-long operations, demonstrating diverse enterprise dynamics. Experiments, including human evaluations, show that it improves performance with less real-world overhead, while also generating internal data as a downstream task and enabling interactions with external contexts.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 12807