The Path to Artificial Superintelligence

Current multi-agent AI systems fail often, performing worse than a single agent because they can’t share goals or think as a group. To fix this structural problem, researchers suggest a two-layer answer with a connection base and a semantic ‘Internet of Cognition’. This new setup lets agent teams work together on new challenges by sharing their purpose, background information, and logic.

Key takeaways

• Current multi-agent systems fail 41% to 87% of the time because they don’t reason together.
• The suggested fix is a structural change, shifting from just making single agents bigger to a horizontal ‘Internet of Cognition’.
• A basic ‘Internet of Agents’ layer lets agents find each other across platforms, verify identity, and send messages.
• The semantic layer allows agents to share goals, combine organizational knowledge, and reason as a team.
• Shared thinking creates new dangers, like accidental task handoffs and corrupted memory, which need fresh protections.

The Architectural Gap: From Solo Agents to a Collective ‘Internet of Cognition’

Current multi-agent systems, used in areas like software engineering and drug discovery, fail frequently. One study of seven open-source systems shows failure rates from 41% to about 87%. According to Vijoy Pandey of Outshift by Cisco, “Connected agents manage coordinated action well… What they cannot do is hold a goal in common and reason toward something none of them was trained to solve.”

The Limits of Vertical Scaling

The industry’s drive for vertical scaling has created larger models and more powerful single agents. Yet, enabling problem-solving across separate systems demands a horizontal scale. Pandey says, “The gap is architectural, not a prompting problem. Without the right coordination layer, naive multi-agent setups can perform worse than a single agent.”

A Two-Layer Solution

The proposed structure uses two core layers. The first is a connection layer named the “Internet of Agents.” It lets autonomous agents find each other, confirm identity, and send messages across different domains. A semantic layer called the “Internet of Cognition” sits on top of this.

This semantic layer supports collective work through three main jobs: sharing intent by aligning goals, sharing context by combining institutional knowledge, and sharing reasoning to make group decisions. This framework allows an agent team to autonomously tackle a new problem without human help.

Building the Connective Tissue: Protocols for Shared Intent and Context

A basic connection layer, shown by the open-source project AGNTCY, lets agents from different systems, companies, and platforms find each other, prove identity, and exchange messages with standard protocols. This “Internet of Agents” layer supplies the essential coordination infrastructure.

Shared Intent via Protocols

Shared intent happens through ‘cognition state protocols.’ This semantic handshake lets agents agree on a goal and bargain to reach it. Outshift’s open-source coordination layer, Mycelium, uses these protocols. In internal tests, unstructured agent groups made a decision only one-third of the time across different situations. Using a coordination protocol that forces agents to state a goal, reveal missing data, and settle conflicts before acting boosted that success rate to 93%.

Shared Context via a Fabric

Shared context comes from a ‘cognition fabric,’ a policy-controlled layer that works as a shared organizational memory and communication network. This lets agent knowledge grow over time instead of resetting every session, avoiding “organizational amnesia.” The fabric guarantees the system’s baseline intelligence always rises, helping agents combine knowledge and develop a cumulative understanding.

Enabling Shared Reasoning and Managing New Risks

Shared reasoning uses two types of cognition engines together. Cognitive amplifiers accelerate shared thinking and modeling, while guardrail technologies (GATs) establish security, cost, and compliance boundaries. Humans stay active participants here, making judgment calls the system sends them in real-time, instead of checking results afterward.

New Risks from Shared Cognition

Sharing cognition in multi-agent systems brings fresh dangers. These include accidental task delegations, harmful prompt injections, memory poisoning, and over-privileged agents with access to permissions and data far beyond their needs.

Reducing these risks needs setting up environment-specific controls. These guard against unintended results from agents working at machine speed. This step is essential because agents have human-like traits but operate much faster, making old oversight methods inadequate.