Multi-agent AI systems are one of the most hyped concepts in enterprise technology right now — and one of the most underperforming. A recent study evaluating seven open-source multi-agent frameworks found failure rates ranging from 41% to 87%. The promise of armies of specialized agents collaborating autonomously remains largely unrealized.

Vijoy Pandey, SVP and General Manager of Outshift by Cisco, has a specific diagnosis for why — and a proposed architecture to fix it.

The Problem Isn't Intelligence, It's Coordination

Pandey's framing is pointed: today's agents are capable reasoners but terrible teammates. Using a healthcare analogy, he describes a system with four specialized agents — symptom assessment, scheduling, insurance, and pharmacy — that can each handle their own domain expertly but cannot coordinate patient care without a human making decisions at every handoff.

"The intelligence is already there. What is missing is the connective tissue that turns four strangers into one team."

The fix, Pandey argues, isn't more vertical scaling (bigger models, more compute, more parameters). The next axis of scale is horizontal — connecting agents across domains, companies, and platforms so they can operate on shared goals rather than siloed tasks.

"Without the right coordination layer, naive multi-agent setups can perform worse than a single agent. The gap is architectural, not a prompting problem."

Two Layers: Agents and Cognition

Outshift's proposed architecture has two stacked layers:

1. The Internet of Agents — a connectivity layer built on AGNTCY, an open-source project now housed under the Linux Foundation. AGNTCY allows agents to discover each other, verify identity, and exchange messages through standardized, open protocols — essentially giving agents an address book and a common language.

2. The Internet of Cognition — a semantic layer that sits on top and enables agents to actually think together. This breaks down into three pillars:

  • Shared intent via cognition state protocols: An open-source coordination layer called Mycelium allows agents to declare a goal, surface missing information, and resolve conflicts before acting. Outshift's internal testing across 14 scenarios found that unstructured agent groups reached a decision roughly 33% of the time; adding the coordination protocol raised that to 93%.
  • Shared context via cognition fabric: A policy-governed memory mesh that prevents "organizational amnesia" — agent insights compound over sessions rather than resetting each time.
  • Shared reasoning via cognitive amplifiers and guardrail technologies (GATs): Amplifiers speed up collective modeling; GATs enforce security, cost, and compliance constraints. Critically, humans are embedded as active contributors at this layer, not passive reviewers of outputs.

The Security Problem No One Is Talking About Enough

Connecting agents at this level introduces serious new risks: unintended permission delegation, prompt injection, memory poisoning, and over-privileged agents with access well beyond their assigned tasks.

Pandey notes that roughly 90% of agents currently have no mechanism to confirm they're even authorized for the task they've been handed — a structural gap that existing enterprise security infrastructure wasn't built to address.

"Agents have human-like attributes but operate at machine speed and scale. Everything we built for twenty years — access control, identity, compliance — was built for humans or machines, not both."

Outshift's response is CASA (Continuous Agent Semantic Authorization), an open-source reference implementation that continuously checks each tool request an agent makes against the original task it was authorized for. In the healthcare example: an agent told to summarize a patient record that begins querying an entire database gets blocked, because the request no longer matches the scoped task.

What This Means for Enterprises Building Now

For founders and technical leaders, Pandey's practical advice is to start experimenting with one cross-functional workflow spanning three or four teams — ideally one that currently requires a human to authorize handoffs between them.

"Stand it up as a small multi-agent system on open, interoperable infrastructure, with a measurable baseline. Track where one agent's insight made another agent better — that is the signal the horizontal axis is working."

The broader competitive context matters here. Google, Microsoft, Anthropic, and others are all racing to define multi-agent coordination standards — Google with Agent2Agent (A2A), Anthropic with MCP. Outshift's bet is on open, interoperable infrastructure under neutral governance (Linux Foundation) rather than proprietary ecosystems. Whether that openness becomes an advantage or a barrier to adoption will depend heavily on enterprise uptake in the next 12–18 months.

The infrastructure is still being written. That's both the risk and the opportunity.