S-Agents: Self-organizing Agents in Open-ended Environments

TL;DR

This work introduces S-Agents, a self-organizing multi-agent framework for open-ended environments, implemented in Minecraft. It combines a directed tree of agents for centralized leadership, an hourglass architecture to unify perception and planning, and a non-obstructive asynchronous collaboration scheme to avoid slowest-agent bottlenecks. The approach advances autonomous coordination without human input and demonstrates superior efficiency in tasks like resource collection and shelter construction. By revealing anthropomorphic organizational behaviors and providing a scalable blueprint for embodied multi-agent systems, the paper contributes to understanding and designing large-scale agent organizations for complex environments.

Abstract

Leveraging large language models (LLMs), autonomous agents have significantly improved, gaining the ability to handle a variety of tasks. In open-ended settings, optimizing collaboration for efficiency and effectiveness demands flexible adjustments. Despite this, current research mainly emphasizes fixed, task-oriented workflows and overlooks agent-centric organizational structures. Drawing inspiration from human organizational behavior, we introduce a self-organizing agent system (S-Agents) with a "tree of agents" structure for dynamic workflow, an "hourglass agent architecture" for balancing information priorities, and a "non-obstructive collaboration" method to allow asynchronous task execution among agents. This structure can autonomously coordinate a group of agents, efficiently addressing the challenges of open and dynamic environments without human intervention. Our experiments demonstrate that S-Agents proficiently execute collaborative building tasks and resource collection in the Minecraft environment, validating their effectiveness.

Figures (11)

• Figure 1: Agent organization in open-ended environments. Agent organization is a group of agents with a certain structure cooperating for shared goals. (1-3) depicts a group of agents collecting scattered rocks; (4-8) illustrates a group of agents building a shelter together. During their collaborative process, they autonomously orchestrated workflows without fixed steps by humans.
• Figure 2: Schematic organizational structure comparison.(a) Solo agent wang2023voyager: Direct interaction with the physical environment; (b) Chain of agents qian2023communicativehong2023metagpt: Specialized agents sequentially perform their designated tasks and command the actions of the next agent; (c) Graph of agents park2023generative: Decentralized structure allowing all agents to command each other; (d) Tree of agents: Centralized structure retaining one agent as a leadership agent (root agent $a_{r}$), with other executor agents (leaf agent $a_{l1}, a_{l2}$) executing commands.
• Figure 3: An illustration of hourglass agent architecture.(a) Hourglass agent framework:The upper segment: Processes inputs like perception and the previous plan. These inputs undergo a series of operations, converging towards a unified and consistent objective (the bottleneck of the hourglass). The lower segment: Involves the decomposition of an objective through hierarchical planning. (b) Progress monitor: Utilizes LLM to assess the current progress status of the ongoing task. (c) Hierarchical planning: Comprises two stages: Task planner and action planner. See Appendix \ref{['appendix:planning_process']} and \ref{['sec:full_prompt_design']} for example of planning and full prompt, respectively
• Figure 4: Comparison of collaboration strategies. (a) involves one agent sequentially executing tasks after another, with no parallelization; (b) is round-based, executing round by round, while (c) is asynchronous. Colored regions indicate tasks being performed, and white regions denote agent idleness.
• Figure 5: Collective shelter construction. The root agent (a1) systematically arranges the tasks and schedules the leaf agents (a2 & a3) for phased execution. More screenshots can be referenced in Appendix \ref{['appendix:execution_process']}.