AGENTHIVE: A Composable Multi-Agent Framework with First-Class Delegation

ACL ARR 2026 January Submission8004 Authors

Keywords: multi-agent systems, framework, delegation

Abstract:

Large-language-model (LLM) agents excel at reasoning and tool use, yet existing multi-agent systems (MAS) rely on static, hand-crafted
topologies and struggle with open-ended, evolving tasks. We introduce AGENTHIVE, a delegation-centric MAS framework that treats
delegation as a first-class primitive: any agent may spawn and coordinate sub-agents, enabling decentralized control without a cen-
tral orchestrator. Through recursive delegation, AGENTHIVE dynamically forms task-adaptive structures—trees, forests, and star
topologies—without predefined agent graphs. We evaluate AGENTHIVE across four real-world domains, showing that MAS emerge au-
tomatically from task demands. The performance gains arise from an expressive, adaptive MAS that allows agents to explore more
deeply and cover a wider solution space. Our results demonstrate that first-class delegation is a powerful paradigm for scaling LLM-based
autonomous systems. An anonymized version of the source code is available at https://anonymous.4open.science/r/anonymous-B-3F03/.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Multi-agent systems, Language models, Agent architectures

Contribution Types: Publicly available software and/or pre-trained models

Languages Studied: None

Submission Number: 8004