Multi-Agent Collaborative Planning

As AI agents get more powerful, it is also possible that they could get more specialized. For instance, in a software development setting, we may create agents that are particularly good at overall software architecture, others that are good at generating code for well-defined functions, others that are good at scanning for bugs or vulnerabilities, and others that are good at fixing them. The potential benefits of these multi-agent systems include achieving higher performance for efficient specialized models than would be possible for more generalist models, and easier engineering or modularization in the case that individual agents are given access to particular task-specific tools or knowledge.

In this work, we propose to develop multi-agent LLM-based approaches to collaboratively curate plans that would be useful to humans, for tasks that involve (1) some coordination between the agents (2) multiple persona-conditioned expert agents and (3) knowledge-backed expertise.

Example tasks include planning corporate events, humanitarian mission planning, carrying out complex in-browser tasks, software architecture and engineering, and resource-constrained coordinated navigation. Prior work has not rigorously evaluated on complex knowledge-conditioned tasks, demonstrated robust multi-agent interactions, or evaluated the trade-offs between a single monolithic agent vs separate modular experts.