CoWork-X: Real-time Multi-Agent Collaboration

A new framework, called CoWork-X, addresses the challenges of collaboration between multiple agents driven by large language models (LLM) in interactive environments. The goal is to enable real-time coordination and continuous adaptation while maintaining a limited token budget.

CoWork-X is based on a co-evolution approach that optimizes collaboration between agents through successive episodes. The system uses a Skill-Agent, which executes tasks by retrieving skills from a structured library, and a post-episode Co-Optimizer, which consolidates learned skills with explicit budget constraints.

Experimental results in simulated environments demonstrate that CoWork-X achieves stable and cumulative performance gains while reducing online latency and token usage. This approach promises to improve the efficiency and effectiveness of multi-agent systems in complex scenarios.