Cisco's Outshift builds 'Internet of Cognition' for AI agents to coordinate without humans

Outshift by Cisco, led by senior vice president and general manager Vijoy Pandey, has identified a critical gap in current multi-agent AI systems: while individual agents can be powerful reasoning engines, they lack the "connective tissue" to work as a coordinated team. Consider a healthcare system with four agents—one handling symptom assessment, another scheduling, a third insurance, and a fourth pharmacy. Today they can exchange data but cannot coordinate patient care without a human making decisions. The reason, Pandey explains, is that "the intelligence is already there. What is missing is the connective tissue that turns four strangers into one team."

Outshift has built that connective tissue in layers. At the foundation is AGNTCY, an open-source connectivity layer now managed by the Linux Foundation. AGNTCY allows autonomous agents across different systems, companies, and platforms to discover one another, prove their identity, and exchange messages using open, standardized protocols. On top of this sits the "Internet of Cognition"—a semantic layer enabling agents to align on shared goals (shared intent), pool institutional knowledge (shared context), and reason collectively (shared reasoning).

The semantic layer rests on three pillars. First, shared intent is enabled through cognition state protocols—semantic handshakes that allow agents to agree on a goal before acting and negotiate toward it. Outshift created Mycelium, an open-source coordination layer organizations can deploy against their own agents. In internal testing, unstructured agent groups reached a decision about one-third of the time across 14 scenarios; with Mycelium requiring agents to declare a goal, surface missing information, and resolve conflicts before acting, success rose to 93%. Second, shared context flows through a "cognition fabric"—a shared institutional memory and communication mesh that prevents "organizational amnesia" and ensures agent insights compound over time rather than resetting each session. Third, shared reasoning is enabled through cognitive amplifiers (which speed up collaborative modeling) and guardrail technologies that create security, cost, and compliance frameworks, with humans actively contributing judgment calls.

Outshift has also released Continuous Agent Semantic Authorization (CASA), an open-source guardrail technology that ensures agent actions remain aligned with the user's original goal. In a healthcare example, an agent authorized to summarize a patient record might begin querying an entire database; CASA would deny that call because the request no longer matches the authorized task. Pandey notes that today's access controls are scoped to a role or session, meaning an agent granted a tool can use it for anything—and "roughly 90% of the time, an agent has no way to confirm it is even cleared for the job it was handed." Coordinated multi-agent systems introduce new risks, including unintended delegations, malicious prompt injections, memory poisoning, and over-privileged agents; CASA and environment-specific controls are designed to mitigate these threats.

Outshift advises organizations to begin experimenting now with one cross-functional workflow spanning three or four teams that currently requires human authorization. By standing up a small multi-agent system on open, interoperable infrastructure with a measurable baseline, and then tracking where one agent's insight made another agent better, enterprises can validate whether the horizontal scaling of intelligence is working. Pandey stresses that "the problems are open, and the infrastructure is still being written," making this the moment for organizations to build on open standards before the landscape solidifies.

The AI industry has long pursued vertical scaling—larger models trained on more data and compute—which has produced reasoning capabilities that function as a "brain" for individual agents. However, multi-agent systems operating across different domains, companies, and platforms have so far failed at rates between 41% and 87%, according to one study of seven open-source systems. The bottleneck is not prompting or model capability but architecture: without the right coordination layer, naive multi-agent setups can underperform a single agent. Outshift's thesis is that the next axis of scale must be horizontal—enabling agents to converge on new problems, negotiate shared goals, and compound institutional knowledge without human intermediaries stitching the seams.

Vijoy Pandey draws a historical parallel: just as humans spent hundreds of thousands of years accumulating individual knowledge that died with each person, then around 70,000 years ago learned to share intent and reason collectively (leading to civilization), AI agents now stand at an analogous threshold. They have been given "silicon genius" and agency, but lack the connective tissue for collective action. Outshift's "Internet of Cognition" framework addresses this by layering semantic coordination (shared intent, context, and reasoning) on top of a connectivity foundation (AGNTCY) that allows agents to discover and authorize one another. The practical implication is that coordinated agents could unlock workflows in healthcare, drug discovery, and software engineering where multiple specialized systems must act as one team—but only if the architectural foundation is solid.

No comments yet. Be the first to share your thoughts!