AI Agents Can Connect But They Can't Think Together

AI agents can be stitched into workflows. They can plug into supervisor models. They can pass outputs to one another in sequence. What they cannot do is share context, intent, or cognitive state in real time — which means every handoff between agents is, effectively, a cold start.

That gap is the next major bottleneck in AI infrastructure. Vijoy Pandey, SVP and GM of Outshift by Cisco, has a precise way of framing it: "Connection is not cognition." His team is building the protocols and infrastructure to close that gap — what Pandey calls the "internet of cognition."

Why Connected Agents Still Can't Think Together

The current architecture of multi-agent AI systems roughly works as follows: individual agents are trained or prompted for specific tasks, then orchestrated by a workflow or supervisor model that routes outputs among them. Each agent operates within its own context window. When an agent receives input from another, it processes that input as new information — it does not inherit the cognitive state, semantic understanding, or accumulated context of the agent that produced it.

The result is what Pandey describes as a "horizontal distributed assistance problem." Agents are assisting each other sequentially rather than reasoning collectively. There is no shared intent, no synchronized context, and no mechanism for agents to coordinate, negotiate, or discover information together without human mediation.

Pandey draws a direct analogy to human cognitive evolution. Humans first became individually intelligent, then developed basic communication through gestures and drawings, and then gradually built collective intelligence — shared intent, coordinated action, the ability to solve problems that no individual could solve alone. That progression took hundreds of thousands of years. His argument is that AI is now on the same trajectory, compressed into a much shorter timeframe, and that the infrastructure for shared cognition does not yet exist.

The Three Protocols Cisco Is Building

Outshift's approach to shared cognition centers on three new protocols, each addressing a different layer of the agent communication problem.

Semantic State Transfer Protocol (SSTP) operates at the language level. It handles semantic communication between agents — analyzing meaning so that systems can infer the appropriate tool or task without requiring explicit instruction at every step. Rather than passing raw outputs, SSTP enables agents to transfer semantic intent, allowing a receiving agent to understand not just what was produced but what it means and what should follow. Pandey's team recently collaborated with MIT on a related framework called the Ripple Effect Protocol, which extends this semantic alignment across cascading agent interactions.

Latent Space Transfer Protocol (LSTP) operates at a deeper level — transferring an agent's internal representational state directly to another agent. In practical terms, this means transferring the KV cache (the key-value attention cache that captures what a model has processed and how it has weighted that information) rather than converting that state back into natural language and re-ingesting it on the other side. The efficiency argument is significant: tokenizing internal state into language, transmitting it, and re-encoding it on receipt introduces computational overhead and information loss at every step. LSTP bypasses that cycle by treating cognitive state as a transferable artifact rather than a language artifact.

Compressed State Transfer Protocol (CSTP) handles the compression problem — how to transfer large amounts of cognitive state accurately and efficiently, particularly in edge deployments where bandwidth and compute are constrained. CSTP grounds only the targeted variants of state information while compressing everything else, enabling the transmission of meaningful cognitive context without transmitting the full state of a large model.

Together, the three protocols are designed to function as a fabric — Pandey's term — that synchronizes cognition states across distributed endpoints. On top of that fabric, Outshift is developing what it calls "cognition engines": systems that provide guardrails for agent behavior and accelerate coordination between agents operating within the shared cognitive infrastructure.

The three-layer architecture Pandey describes: protocols, fabric, and cognition engines. Each layer is a prerequisite for the next.

What Cisco Has Already Built — and What It Measured

Alongside this longer-horizon infrastructure work, Pandey described a concrete internal deployment that illustrates what current-generation multi-agent systems can achieve when properly implemented.

Cisco's site reliability engineering team was facing a scaling problem: increasing product output and code volume without a commensurate increase in team size. Pandey's team deployed more than 20 AI agents — a mix of in-house and third-party — with access to over 100 tools via the Model Context Protocol (MCP), which is integrated into Cisco's security platforms.

The agents automated more than a dozen end-to-end workflows, including CI/CD pipelines, EC2 instance provisioning, and Kubernetes cluster deployments. The measured outcomes: certain deployments dropped from hours to seconds, and agents reduced 80% of the issues the SRE team encountered in Kubernetes workflows.

Pandey is careful to contextualize this: "It does not mean that I have a new hammer and I'm just gonna go around looking for nails. You still have deterministic code. You need to marry these two worlds to get the best outcome." The deployment works because it matches AI agents to tasks where non-deterministic reasoning adds value, while keeping deterministic code in place where predictability is required.

Separately, Pandey noted that Cisco has used AI to improve error detection in large networks from 10% to 100% — a result that speaks to the diagnostic and pattern-recognition strengths of current-generation models when applied to well-defined operational problems.

The Open Infrastructure Requirement

Pandey is explicit that the internet of cognition cannot be a proprietary system. For distributed shared intelligence to function at scale, the protocols and fabric must be open and interoperable — accessible across vendors, platforms, and agent frameworks. Cisco's open-source project Agntcy addresses the discovery, identity, and access management, observability, and evaluation layers of this infrastructure.

The analogy to the early internet is deliberate and precise. TCP/IP did not belong to any single company. HTTP did not belong to any single company. The protocols that enabled the internet's scale were open standards that any system could implement. Pandey's argument is that shared agent cognition requires the same approach — proprietary solutions will fragment the ecosystem before it can reach the scale at which collective intelligence becomes genuinely useful.

What This Means for Teams Building on AI Today

For technical leaders evaluating AI agent infrastructure, Pandey's framework offers a useful diagnostic. The question is not whether your agents are connected — most modern agent frameworks handle orchestration reasonably well. The question is whether they share context meaningfully across sessions and handoffs, or whether each agent interaction is effectively a new conversation with no accumulated understanding.

For most current deployments, the honest answer is the latter. The protocols Outshift is building — SSTP, LSTP, CSTP — represent the infrastructure layer that would change that. They do not exist at a production scale yet. But the direction they point is toward agent systems that reason collectively rather than sequentially, which is a qualitatively different capability from what enterprise AI deployments can currently achieve.

Understanding where the infrastructure is today — and where it is going — is part of building an AI strategy that holds up over time rather than optimizing for what is available right now. At Winsome Marketing, our AI strategy and integration work is built on that longer view. If you want to think through what this infrastructure trajectory means for your business, our team is ready to help.