AI multi-agent systems need a shared cognition layer to coordinate across domains
If realized, a shared cognition layer would let agents align goals, pool memory, and reason together across different systems, enabling coordinated problem solving at scale beyond what any single agent can do today.
At a glance
- AGNTCY is an open-source connectivity layer for agents, now under the Linux Foundation.
- Mycelium is a coordination layer that helps agents declare goals, surface missing information, and resolve conflicts; internal testing raised decision-making success from ~33% to 93% across 14 scenarios.
- The approach rests on three pillars: shared intent through cognition state protocols, shared context via cognition fabric, and shared reasoning to trade off between options.
The story
MIT Technology Review describes a path toward artificial superintelligence built from coordinated AI agents rather than a single monolithic system. The piece outlines a horizontal architecture where multiple domain experts—each an AI agent—can share intent, context, and reasoning to tackle wide-scale problems that no individual agent was trained to solve. The concept hinges on a semantic layer called the Internet of Cognition, paired with a connectivity layer the article calls the Internet of Agents, enabling autonomous agents to discover one another, prove identity, and exchange messages across domains.
Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, argues that the missing piece is architectural rather than a prompting fix: the right coordination layer is needed to turn scattered agents into a coordinated team. He says the next axis of scale for AI is horizontal, not just adding more data or bigger models.
Outshift has built AGNTCY, an open-source connectivity layer intended to let agents across different systems, companies, and platforms find and talk to one another using standardized protocols. The goal is to enable the Internet of Cognition to recruit shared goals, context, and collective reasoning among diverse agents.
The article notes that multi-agent systems are already explored in software engineering, drug discovery, and scientific simulations, but current performance has been uneven. A cited study reports a failure rate between 41% and about 87% when evaluating seven open-source multi-agent systems, highlighting a gap that is described as architectural rather than a prompting problem.
Pandey emphasizes that without a proper coordination layer, naive multi-agent setups can underperform compared with a single agent, and that the real step is a team of agents converging on a new problem without human stitching.
Three pillars anchor the proposed stack: cognition state protocols for shared intent, cognition fabric for shared context, and shared reasoning for collective trade-offs. The article explains how these layers would enable agents to align goals, pool institutional memory, and negotiate outcomes.
The piece also highlights Mycelium as an open-source coordination layer that organizations can clone to run against their own agents. It cites internal testing results showing substantial improvements in coordination and decision-making when a formal protocol is used.