Title: Is It Possible for LLM Agents to Maintain Long-Term Organizational Dynamics?

Abstract: While large language agents are seeing growing adoption in social simulations, their capacity to uphold consistent behavior within structured organizations remains an open question. Such environments require goals to cascade down hierarchies, tasks to rely on the completion of preceding steps, and artifacts to accumulate over extended periods. To address this, we frame long-horizon organizational simulation as a coordination challenge centered on memory. We propose TaskWeave, a hierarchical framework that preserves planning states via a Formulate-Partition-Diagnose-Align cycle and ensures execution grounding through trace memory that is aware of dependencies. Our evaluation of TaskWeave involves a year-long simulation of an IT company, benchmarking it against other multi-agent systems on metrics of organizational coherence, execution grounding, and downstream enterprise NLP utility. The results demonstrate that TaskWeave facilitates coherent, long-term organizational dynamics, generates grounded artifacts, and adapts to external environmental shifts. These outcomes indicate that structured simulation memory serves as a critical mechanism for developing dependable LLM-based organizational simulators.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC