Outshift by Cisco proposes open infrastructure to coordinate AI agents into teams

Outshift by Cisco argues that today’s AI agents can exchange data but still lack the connective tissue to coordinate actions without human oversight. The missing layer, per Outshift, is a semantic “Internet of Cognition” that lets agents share intent, context, and reasoning so they can converge on goals neither agent was trained to solve individually.

The company describes a two-layer technical stack: beneath the semantic layer sits an “Internet of Agents” connectivity layer that lets autonomous agents discover each other, prove identity, and exchange messages across systems, companies, and platforms. Together, the layers are positioned as a step toward distributed artificial superintelligence, says Vijoy Pandey, SVP and GM of Outshift.

Outshift has released an open-source connectivity layer called AGNTCY under the Linux Foundation to implement the Internet of Agents. On top of that, Outshift proposes a semantic coordination protocol called Mycelium that lets agents declare goals, surface missing information, and resolve conflicts before acting. In internal tests across 14 scenarios, unstructured groups reached a decision about a third of the time, while groups using the coordination protocol succeeded 93% of the time.

The stack’s third pillar is a shared context layer called a cognition fabric, which acts as a policy-governed memory mesh so agent insights compound over time rather than resetting each session. Outshift calls this an antidote to “organizational amnesia,” ensuring the system’s baseline intelligence only increases.

To manage risk, Outshift introduces Continuous Agent Semantic Authorization (CASA), an open-source reference implementation that enforces task-level authorization. CASA reads an agent’s goal and checks each tool request against that task, denying calls that drift beyond the authorized scope. Pandey notes that today’s controls are scoped to roles or sessions, not tasks, and estimates roughly 90% of the time an agent has no way to confirm it is cleared for the job it was handed.

Pandey frames the progression as analogous to human history: for hundreds of thousands of years humans grew individually smarter, but gains died with each person until about 70,000 years ago when humans learned to share intent, build cumulative knowledge, and reason collectively—becoming civilization. “Agents are at the same threshold,” he says. “We have built the silicon geniuses and given them agency. What they lack is the layer that let humans go collective.”