Cisco’s Outshift claims its agent coordination layer hits 93% decision success

The multi-agent systems everyone’s excited about have an embarrassing problem: they fail constantly. Vijoy Pandey, SVP of Outshift by Cisco, cites a study showing failure rates between 41% and 87% for open-source setups. The issue isn’t intelligence—it’s architecture. Agents are like brilliant specialists who’ve never met, exchanging data but incapable of reasoning together toward a goal none of them was trained for. “The gap is architectural, not a prompting problem,” Pandey says. “Naive multi-agent setups can perform worse than a single agent.”

Outshift’s answer is a two-layer stack it’s calling the Internet of Cognition. Beneath it sits AGNTCY, an open-source connectivity project now under the Linux Foundation that lets agents discover each other, prove identity, and message across domains. On top of that sits a semantic layer built on what Pandey calls cognition state protocols and cognition fabric. The protocol layer—called Mycelium—forces agents to declare goals, surface missing information, and resolve conflicts before acting. In internal testing across 14 scenarios, unstructured agent groups reached a decision roughly a third of the time. With Mycelium’s coordination protocol layered on, that number jumped to 93%.

That’s a striking delta, and if it holds outside the lab, it addresses the core fragility of multi-agent systems: organizational amnesia. Cognition fabric functions as a shared institutional memory and communication mesh, so insights compound over time rather than resetting each session. The vision is less about making individual agents smarter and more about giving them the connective tissue to think collectively. Pandey draws an analogy to human prehistory—for hundreds of thousands of years, individual intelligence grew but the gains died with each person. Everything changed when humans learned to share intent and build cumulative knowledge.

Skepticism is warranted. A 93% decision rate in controlled scenarios doesn’t guarantee robustness in the wild, and the history of coordination frameworks is littered with protocols that work beautifully on whiteboards and crumble under real-world messiness. But the architectural diagnosis feels right. The industry has obsessed over vertical scaling—bigger models, more compute—while neglecting the horizontal problem of making agents interoperate intelligently. If Outshift’s open-source bet pays off, it could shift the conversation from building better brains to building better teams.