Autonomous AI agents equipped with coordination mechanisms reached alignment in approximately 93% of test cases, up from roughly 36% without those tools, according to experiments conducted by Outshift, Cisco's emerging-technology incubation group. The findings underscore a challenge emerging as enterprises deploy multi-agent systems: connecting agents doesn't guarantee they'll actually work together. Outshift is now proposing what it calls an Internet of Cognition, an architecture designed to help agents move beyond information exchange toward shared intent, context, and coordinated action.
The alignment problem becomes more acute as agent populations grow. Recent research presented at ACL 2026, a benchmark called SILO-BENCH, identified what researchers termed a "Communication-Reasoning Gap": agents communicated actively but frequently failed to convert that interaction into effective collective computation. On the most complex tasks, performance deteriorated sharply as the number of agents increased. A separate 2025 NAACL study found that agents handled decisions based on environmental information reasonably well, but struggled when successful coordination required grasping another agent's beliefs and intentions. Outshift has been testing an open-source project called Mycelium to address these failures, and the company says the magnitude of improvement in its internal experiments illustrates a problem enterprises will soon confront at scale.
Papi Menon, Chief Product Officer and VP of Product Management for Outshift, told the publication that the team is "considerably further into autonomous multi-agent workflows than many enterprises" and has begun encountering problems others haven't yet reached. "Agents don't align and they kind of tend to diverge," Menon said. The coordination layer matters for another reason, according to the report: humans can't exercise judgment over systems they don't understand. Menon described the challenge by noting that even as agents become more autonomous, their activity ultimately has to be "explainable, rationalizable and justifiable to a human."
The distinction between communication and coordination sits at the heart of what Outshift is attempting to solve. Agents may be technically capable of exchanging data while still lacking a shared understanding of what that data means, what they're collectively trying to accomplish, or how to resolve competing interpretations. The report offers a concrete example: the phrase "Priority 1" might mean an incident requiring resolution within an hour to one organization's IT agent, but could imply an immediate life-or-death situation to an agent operating in a healthcare environment. Both agents received identical words but didn't receive the same meaning. As agents increasingly operate across systems, functions, and eventually enterprises, those semantic differences become operational problems that connectivity alone can't fix.
Outshift's proposed architecture focuses on three areas: protocols that establish shared intent and coordination, a cognition fabric that supports shared and policy-governed context and memory, and cognition engines that can help agents negotiate, coordinate, or enforce guardrails. The company says it's already seeing this coordination layer work in practice. Menon described work between Outshift and ServiceNow in which agents operating on separate platforms participated in the same end-to-end enterprise IT workflow, connected in a matter of days rather than requiring substantial custom integration. But the real test comes when workflows contain not two relatively bounded agents, but 20 autonomous agents representing security, finance, procurement, HR, suppliers, and customers, each with different data, goals, permissions, and interpretations. At that scale, the report argues, counting how many agents an enterprise has deployed will become an increasingly meaningless measure of AI maturity. A company could have 500 highly capable agents and still have very little collective intelligence. For decision-makers wondering whether their agents truly collaborate, the question becomes less about scale and more about whether agents can think together without losing context, accountability, or judgment along the way.