Author(s): Aniket Sanyal

Originally published on Towards AI.

Multi-Agent LLMs, MACE, and the Hidden Collaboration Problem

A new AI research paper shows that agentic AI systems can look collaborative while quietly failing at one of the most basic team skills: learning who to ask.

After the introduction, the article explains that multi-agent teams can “sound busy” yet fail to explore—agents often prematurely commit to one teammate, over-trust early signals, ignore other potentially better collaborators, and repeat interaction habits even when evidence is weak. It frames the core issue as an exploration-and-learning problem (not just conversation design), and introduces MACE (Multi-Agent Contextual Exploration) as a structured decision layer that encourages under-tested peers based on uncertainty and contextual usefulness—outperforming prompt-only approaches that can still under-explore (even worse than random peer selection in some settings). The piece discusses how MACE helps reduce “regret” (missed opportunity from poor collaboration paths), can transfer learned interaction strategies to new benchmarks, and that simply using bigger or stronger models does not solve the coordination gap—agent orchestration must be explicitly engineered, logged, and measured, especially in products where teams combine diverse agents, tools, and knowledge sources.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI