Can AI Agents Learn to Think Together? One Cisco Team Thinks It Has the Connective Tissue

Key points

• Outshift by Cisco, a research unit inside Cisco, has released two open-source projects, AGNTCY and Mycelium, designed to let separate AI agents coordinate without human supervision.
• Internal testing found uncoordinated groups of AI agents reached a correct decision roughly one-third of the time; adding a goal-alignment protocol pushed that rate to 93 percent.
• One study cited by the team recorded failure rates of between 41 percent and 87 percent across seven open-source multi-agent systems, the technology that chains several AI agents together on a single problem.
• A new security tool called CASA (Continuous Agent Semantic Authorization) checks, in real time, whether each action an agent takes still matches the task it was originally given.
• Both AGNTCY and Mycelium are now managed under the Linux Foundation, the non-profit that oversees widely-used open-source software.

Picture a hospital with four separate AI agents: one assesses symptoms, one handles scheduling, one talks to insurers, and one coordinates the pharmacy. Each is sharp in its own lane. None of them can agree on a treatment plan unless a human steps in to translate between them.

That handoff problem, says Vijoy Pandey, the senior vice president running Outshift by Cisco, is the core unsolved puzzle in modern AI. The technology can reason. It just cannot collaborate.

Why is coordinating AI agents so hard?

Today's AI systems have mostly grown by getting bigger: larger models, more training data, more computing power. That approach produces smarter individual agents, but size alone does not teach them to work as a team.

Connected agents handle tasks well when the task looks familiar. What they cannot do, Pandey argues, is hold a shared goal and reason toward a solution none of them was designed for. That is an architectural gap, not something you fix by writing better instructions.

Research backs the frustration. One study of seven open-source multi-agent systems found failure rates ranging from 41 percent to nearly 87 percent. A single well-designed agent often outperforms a poorly coordinated group.

What has Outshift actually built?

Two tools, both free to use and modify under open-source licences.

AGNTCY is the networking layer. Think of it as a directory and postal service for AI agents: it lets agents from different companies and platforms find each other, verify who they are, and exchange messages through shared standards. It now lives under the Linux Foundation.

Mycelium sits on top. It is a coordination protocol that forces agents to declare a common goal, flag missing information, and settle conflicts before anyone acts. In Outshift's own tests across 14 scenarios, unstructured agent groups reached a correct decision about one-third of the time. Mycelium brought that to 93 percent.

Pandey calls the combined result an "Internet of Cognition," a shared layer for intent, memory, and reasoning that he compares, with some ambition, to the cognitive leap that turned scattered early humans into cooperating societies.

Should people worry about AI agents acting unsupervised?

Yes, and Outshift says it does too. Linking agents together creates new risks: one agent can accidentally grant another permissions it should not have, bad actors can inject false instructions into shared memory, and agents can take actions far outside the scope of their original task.

The answer Outshift proposes is CASA, Continuous Agent Semantic Authorization. The tool reads what an agent is trying to accomplish, then checks every individual action against that stated purpose in real time. In a healthcare example Pandey gives, an agent asked to summarise a single patient record should not be querying an entire hospital database. CASA would block that call.

"Today's controls are scoped to a role or a session, not to the task," Pandey told MIT Technology Review, which first covered the project. "Roughly 90 percent of the time, an agent has no way to confirm it is even cleared for the job it was handed."

What does this mean for businesses or patients today?

Neither AGNTCY nor Mycelium is a finished product a company can switch on tomorrow. They are building blocks, and the standards around them are still being written.

Pandey's practical advice for businesses curious about multi-agent systems: pick one cross-functional workflow that currently needs a human to approve each handoff between three or four teams, then run a small, measurable pilot on open infrastructure. Watch for moments when one agent's output makes another agent perform better. That, he says, is the signal that horizontal coordination is actually working.

For patients or anyone on the receiving end of AI-assisted services, the honest position is that this technology is not yet running your care. The research goal is to make eventual systems safer and more coherent, not to remove humans from the loop today.

Common questions

What is a multi-agent system?

A multi-agent system is software that uses several separate AI programs, each with its own specialty, to tackle a problem together, the way a relay team hands a baton rather than one runner doing everything.

Is "artificial superintelligence" actually being built here?

No. Pandey uses the phrase to describe a long-term direction, not an imminent product. What Outshift has released are coordination tools for today's narrowly-focused AI agents, not a system that exceeds human intelligence across all tasks.

Who can use AGNTCY and Mycelium?

Both are open-source, meaning any developer or organisation can download, inspect, and adapt the code for free. They are maintained under the Linux Foundation.