Vijoy Pandey, who heads Cisco's incubation group Outshift, argues that the next layer of computing has to carry meaning rather than data: agents that are now connected still fail to think together, and Outshift has released open-source tooling to measure and correct that. He then extends Cisco's "horizontal scaling" thesis to quantum networking, where he reports a record entanglement-swapping result on New York fiber, and ends on the case for hardware that computes on probabilities.
From connected agents to silent failures. Pandey traces agentic computing from isolated agents, to "islands" of sub-agents inside one harness, to agents connected across vendors, clouds and organizations, which Outshift has deployed for 18 months across several verticals. Connection, he says, is not cognition. Agents pass valid messages but mean different things by them: "roll back" to an SRE agent is the last config, to a software-engineering agent the whole version, and "ready" means something different to sales, legal and engineering on a deal. A system can grow more consistent through these failures without growing more correct. Analysing their own deployments plus public ChatDev and MMLU traces, Outshift found at least one cognitive failure in every session, and accuracy of 98%, 83% and 16% at one, two and three failures, all invisible to standard observability.
A stack for shared cognition. He proposes three stages modelled on human teams: align intent and ground terminology; coordinate and negotiate from each agent's local incentives toward a global optimum; then verify the result at the semantic layer. He calls this the "internet of cognition" and frames it as the agent equivalent of language triggering collective intelligence about 70,000 years ago, crediting Michael Tomasello's work. The connectivity layer below it, he says, is converging: AGNTCY, A2A and MCP sit in the Linux Foundation, with common discovery, identity and the SLIM messaging layer.
What ships. The observability and explainability platform OXP reads OpenTelemetry traces and flags intent, coordination or verification drift. Like lane assist, it triggers a protocol only when drift appears: Accord locks a shared intent contract in four phases before work starts, and Concord resolves conflicts during a task. Both are open source and run as a skill, a sidecar, or over A2A, MCP and SLIM. Asked why Cisco funds this, he says Outshift de-risks markets before Cisco builds products, and that open ecosystems are more profitable for every player.
Quantum by networking. Cisco's niche, he says, is horizontal scaling, where others scale algorithms or build bigger machines. Useful quantum computing needs roughly 100,000 to 1 million physical qubits against about 1,000 today, so Cisco aims to network many nodes into one logical machine, routing qubits without measuring them. With Qunnect, NYU and a fiber provider, it ran entanglement swapping in February over 17.6 km of Manhattan–Brooklyn fiber through 60 Hudson, at rates he puts at 5,000 times beyond any lab result.
Probabilistic hardware. Agentic workloads are probabilistic but run on deterministic GPUs, a mismatch he says costs energy, money and size. Between binary and quantum lie classical-physics chips that compute on distributions: Extropic's thermodynamic systems settle into low-energy states, and Unconventional AI's oscillator chips settle into synchronized patterns. Both yield local optima, which he argues suits search problems such as drug discovery, where there is never a single best answer, only the best for now.