Cisco wants two extra network layers so agents can share intent

Nathan Labenz opens with a three-word job description that almost everyone in America has heard of and almost no one in AI can unpack: Cisco moves information around. Vijoy Pandey, SVP and GM of Outshift by Cisco, does not argue with the slogan. He unpacks it into four businesses — networking, security, observability, and collaboration, the last of those being Webex — then reframes the company for an ML audience. Cisco, he says, is a distributed-systems shop. Scaling a server up is one trick. Connecting many servers into a cluster is the other, and that is the one Cisco has spent decades on.

That framing matters because the OSI seven-layer model is the formal map of the world he wants to extend. Physical, link, network, transport, session, presentation, application: the stack everybody pretends to follow and then works around. In practice, he tells Labenz, three names do most of the work. Ethernet lives at physical and link. TCP/IP, layers three and four, is the internet. HTTP and HTTPS are layer-seven application protocols. The rest, he admits with a shrug that will annoy networking purists, is formalism.

How much of that stack already runs on machine learning? Pandey draws a hard line. The data plane — routing tables, forwarding, the path between two people on a call — stays deterministic. "I don't want that to be left to chance." The control plane is different. BGP, outage reroutes, undersea-cable failures, inter-company policy between a Google and an Amazon: those software layers have used ML for years. Anomaly detection, Super Bowl traffic prediction, choke-point failure forecasts, customer sentiment, even Webex noise reduction are old pipelines. What changed is that generative models finally became practical inside the enterprise, and they sit on top of the stack rather than reaching down into the forwarding path.

The first generative-AI system Outshift actually shipped was Project Jarvis, later rolled into Splunk, Webex, and other Cisco business units. The idea is almost blunt: take the agentification already happening in software development and point it at site reliability engineering. SREs spend their days on repetitive work that can be automated; they already treat infrastructure as software. Jarvis is that pipeline with a swarm of agents on it.

The production name is CAIPE, the Community AI Platform Engineer, complete with a superhero-cape logo. It is a multi-agent system, not a chatbot with tools. More than five user interfaces feed it. It makes a hundred-plus tool calls across cloud providers, cloud-native environments, and on-prem, spanning observability, orchestration, networking, and security. It covers more than ten workflows. Roughly twenty agents work together. Load on the team is down 30 percent. Forty percent of the tasks the team used to handle are now end-to-end agentified — "we don't even worry about it." Response time went from hours to instantaneous, and Pandey mentions morale in the same breath as efficiency, because the SRE group is no longer drowning in daily fire drills.

Cisco did not keep the system in-house. It is open source through the Cloud Native Operating Excellence community, with members Pandey lists as "fascinating names": Adobe, AWS, Cisco, Nike, and a broader set of enterprises that have been growing the project. He treats the scoreboard as a proof of plumbing, not as the destination. The destination is an Internet of Cognition — infrastructure so that agents from different vendors, representing different interests, can find one another, share intent, and solve problems that none of them saw in training.

Asked what the frontier of the multi-agent system can handle, Pandey starts with a definition rather than a benchmark. The version of AGI and ASI that interests him is a team of agents collaborating to solve something net new — "completely novel, that has not been in that training data at all" — with no human intervention, 100 percent of the time. He knows he is mixing economic and technical definitions. He is willing to bet on that mix anyway. Timelines for ASI, he notes, keep sliding forward and back. The more important fact is that the whole industry is chasing one vector: vertical scaling. Bigger models, better reasoners, more data, more compute, more parameters.

The second axis is horizontal: can intelligence scale out so the collective is greater than the individual? Cisco's distributed-systems DNA is why he cares. Individual "brains" are getting smarter, but they have not learned to think together — shared intent, shared cognition, collective innovation on problems nobody has seen. "If you want to pull in those timelines, we have to tap into that horizontal axis." Labenz adds a measurement wrinkle from Ajeya Cotra, now at METR. The famous time-horizon graph still asks how long a task would take one human. Up to about sixteen hours, Cotra had written, one person versus a split team is not that different because coordination has a fixed cost. Past a couple of weeks, parallel work wins, so she suggests scoring models against a well-organized human team. Pandey takes the other side of the same coin: that ruler is still one AI, possibly spawning sub-agents. He wants specialists from different vendors and different models, because that is where human teams actually get their strength.

He reaches for human evolution to make the point. Tool use, symbolic communication, painting, raising flags: all of that existed for hundreds of thousands of years while the intelligence curve stayed relatively flat. Language, around 70,000 years ago, was the step function. Suddenly a group could stop charging a hill and instead agree on a strategy, break the job into pieces, assign by expertise, align on intent, coordinate, and execute on a problem they had never seen. The same trajectory, he says, is playing out in silicon. Claude, GPT, and Gemini now spawn sub-agents, "but it's still not. It's like sub-processes within my brain." What is missing is teams of agents that share intent, negotiate, keep institutional knowledge, and innovate on a new problem statement. That network is what he calls the Internet of Cognition, and it will be multi-vendor by definition, because subject-matter experts do not all come from one shop.

Labenz makes the abstract concrete with the episode's own prep. He used Tasklet, running on Claude, to research Pandey, pull publications, cross them against years of Cognitive Revolution outlines, and draft questions. He still rewrites, because an AI draft does not put him in the room. On the other side, someone on Pandey's team — later named as Rebecca — did a similar pass over Outshift's work and Labenz's recent themes. Both packets were useful. Neither agent knew the other existed. The outputs moved human to human by email. Labenz was left to synthesize two ships that had passed in the night.

Pandey likes the example because it is simple enough to see the missing layer. Labenz's agent has a local objective: produce the best episode in the Cognitive Revolution series, using the history of guests and what would be net new for that audience. Outshift's agent has a different context: speaking engagements and what has worked for people who listen to Outshift. The two circles overlap, and the shared goal is a conversation worth the audience's time. The concessions — "that might be a little bit too marketing-oriented... let's get a little real" — are being made by humans. There is no agentification of that pipeline.

The future he wants is hands-off: throw the two agents together, let them converge on common intent with concessions on both sides, and keep the shared context so the next guest, or the next Outshift speaker on a different show, does not start from zero. Today the humans still do intent alignment, coordination, negotiation, and context-building in Google Docs or SharePoint. "It's a human glue that is enabling collective intelligence." The Internet of Cognition is the attempt to turn that glue into software.

Pandey walks through a multi-agent system his team built for a healthcare provider. Patients call in and need to be routed to the right doctor given history, the doctor's availability, and insurance. Four agents sit on the board. Scheduling, on the hospital side, is the chatbot that talks to the patient and books the slot. Insurance is a third-party payer agent. Diagnostics is third-party. Pharmacy is third-party. Only scheduling belongs to the hospital. The job is straightforward: get the caller to the right provider given symptoms, insurance, and pharmacy record. The four agents, at the start, could not talk to each other at all.

Step one is discovery, connection, identity, access, and stitching them into a multi-agent system. About a year earlier Cisco had launched what it called the Internet of Agents and an open-source project, AGNTCY — spelled A-G-N-T-C-Y, landing page at agntcy.org, now under the Linux Foundation. The project is the basic plumbing: discover the four agents, attach identity and access attributes, let traffic flow, and watch whether they deliver. Agent-to-tool traffic goes over Anthropic's Model Context Protocol. Agent-to-agent traffic goes over Google's A2A. Cisco is a foundational member of both. Observability, including a first pass at evaluation that Pandey says is not fully solved, is the fourth pillar.

That is what is deployed and working. What is not happening is anything above the pipe. The four agents can send messages, but the payload is opaque — "just a blob," a binary object whose contents nobody in the stack understands. Coordination, alignment, shared memory, and shared context sit on the grayed-out right-hand side of his diagram, still done by a human in the loop. Each agent has a local KPI: scheduling is scored on time-to-route, diagnostics on outcome confidence, insurance on ROI. Somebody has to step in and tell them to give a little to get a little, relaxing local metrics so the global outcome for the patient can land. Intent, coordination, and mixed patient-specific plus population knowledge all still require a person in the middle.

Labenz analogizes today's operator to the old telephone switchboard: fetching API keys, hunting for a pharmacy-agent ID, plugging things in by hand. Pandey splits consumer from enterprise. Consumers often pick agents by reputation, pedigree, or even by letting an LLM generate one on the fly. Enterprises have procurement cycles, customer trust, and responsible-AI review; they do not work that way. AGNTCY's answer is four software entities with APIs: discovery, identity and access, communication, and observability.

Discovery is an agent directory, searchable by capability, closer to DNS than to a fixed URI. You can register agents, tools, nested multi-agent systems — "turtles all the way down" — plus tool and data-source access. A lookup may return a remote API endpoint or a Git branch you pull and run locally. The backend is a distributed hash table, built to search across huge numbers of entries with nobody owning the catalog. After discovery comes identity. Agents have human-like semantic, non-deterministic communication and machine speed and scale, so classic role-based access is the wrong instrument. Communication then splits: MCP to deterministic tools and data sources, A2A between agents. Observability is not just whether the container is up. "Your service might be up, but the agent might be misbehaving." Cisco and Microsoft have been pushing OpenTelemetry extensions for that agentic layer.

Labenz presses on the internet's drift toward walled gardens and app stores — OpenAI trying, in his phrasing, to be a curated store with reviews and ads. Pandey agrees that directory and identity are "the source of everything else that follows." Control those, and you control reputation, security, pricing, and commerce. That is why AGNTCY put both on DHTs. Nobody owns the directory; participants join a decentralized infrastructure. Identity can plug into Okta, Duo, and other familiar IDPs, or into a decentralized identity provider someone deploys themselves. Enterprises may start with Microsoft Active Directory because that is what they know, then move toward task- or tool-based schemes. The architecture leaves the door open rather than forcing a single owner.

Identity, Pandey says, is the biggest hurdle to deploying as much autonomy as enterprises will tolerate. Agents talk in natural language, make decisions, and call tools. Credit-card access for a consumer agent already makes him careful; enterprises have to be surer. Existing systems were built for humans. Role-based access control, RBAC, is "a very human-oriented view." When machines did not fit, the industry added attribute-based access control, ABAC — "almost like saying RBAC but applied to machines," because ordinary software does not change persona. Agents do. They might be Vijoy this hour and Nathan the next.

So he goes back to primitives. When Cisco gives Vijoy access, it really grants tools, tasks, and transactions. AGNTCY's proposal is TBAC: tool, task, and transaction-based access control. Default to the lowest layer of access. Elevate only for the tool, task, or transaction at hand, then snap back. Do it ephemerally; do not give agents longevity of privilege. The catch is that you cannot name the task unless you can read the conversation. Deterministic software announced what it was doing. Semantic agents do not. To grant TBAC you have to tap layer nine — the semantic flow between humans and agents, or agent and agent — and infer the job before you open the door. Least privilege, in this design, depends on a cognition-state layer that does not exist yet.

That is the motive for extending OSI. Today's endpoints were deterministic and exchanged data. Tomorrow's endpoints are probabilistic intelligence, exchanging cognition state among a person and a handful of agents on each side. Without structure, you cannot tell whether a dialogue is in discovery, intent alignment, coordination, negotiation, or execution. OpenClaw made the failure mode vivid: a cluster of agents talking NLP, prompted by humans, "not building consistent state," all over the map. Enterprises that want convergence and constructive emergence need more than the wild west of unconstrained language. Layer 8 is syntax: an agent built on Vertex and one built on LangGraph or Bedrock do not share grammar or payload format; MCP, A2A, and the AGNTCY architecture live here. Layer 9 is semantics, looking inside the packet, wrapping a header that says "I am discovering," "I am negotiating," or "I am about to call a tool," so trust and governance have something less than fully probabilistic to hold onto.

The Internet of Cognition, on Pandey's slide, has three layers. The protocol layer is how intelligence endpoints talk. He names three classes. Semantic state transfer is natural language, the obvious baseline. Latent-space transfer ships the entire KV cache from one model to another — "like implanting a Neuralink chip" so neither side pays tokenize, transmit, and detokenize. Closed-weight providers will not allow that implant today; open-weight models might. Compressed state transfer sits in between, for edge cases: a phone, a MacBook, a wearable talking to something in a data center. All three are still that protocol layer.

Above protocols sits a cognition fabric: many-to-many, real-time, semantic, with a memory slot you fill yourself. Cisco does not dictate the store. Plug in open-source memory such as Mem0, or BigQuery, or whatever you already run. What they do suggest is storing mixed kinds of shared state — ontologies, beliefs, working memory, knowledge graphs — so the cluster has common context. The third layer is cognition engines, which he hangs on a Raj Reddy line: AI should be a cognitive accelerator or a guardian angel. Those engines are meant to be mostly transparent, like the note-taker in a meeting who says "don't mind me." Summarization, compliance, security: they keep collective intelligence inside the enterprise's swim lanes.

Labenz puts the safety objection on the table. Should we let AIs talk in languages we cannot read? Chain of thought is still mostly vanilla, OpenAI has said, but a Meta paper on thinking in latent space found pathfinding worked better in superposition, and ChatJimmy.ai has burned a model into silicon at 15,000 tokens per second — 2,000 tokens back in a sixth of a second. Pandey's answer runs the other way from many safety arguments. He wants emergent behavior. Diversity of models and vendors is a reliability practice even in deterministic software, where identical code shares the same zero-days. More motion, more chaos, more innovation — and then the engines as laws of physics that say you cannot walk through the table. "A human cannot make an API call. A human doesn't have a KV cache that I can send across." Agents can. Use those properties. The failure modes are divergence, siloed over-specialization, and the rest, which is why the engines watch for those too, and why he wants game-theoretic models of multi-agent competition borrowed from decades of human-behavior research.

OpenClaw is already a preview, even if it is still chatter. Peter's `soul.md` gives an agent a personality close to his own; Pandey's friends have customized theirs. Play that forward a year or two and you have agents with expertise and personality sitting on human teams. The infrastructure question is how those mixed societies share intent, coordinate, and stay inside enterprise guards. "I don't think it's science fiction. You can see that path taking shape."

Labenz closes on a line he repeats across episodes: anything too concentrated becomes dangerous, and concentrated intelligence looks like the same pattern. The alternative has to be distributed, buffered, networked, more ecological, and permissionless on at least some margin. Pandey flips the worry. In the enterprise both of Labenz's concerns are requirements, not taste. Companies will not deploy solutions whose guardrails are not well thought out; that hurdle is what keeps the generative-AI train from running away inside a bank or a hospital. Adoption itself is the governor. And no enterprise will be a single-vendor agent shop. Look around, he says: ServiceNow, Salesforce, Microsoft, OpenAI, Anthropic, Google, Cisco — those agents will all be in the building, and they will have to connect, share cognition, and align on intent, because one agent will not do every job. "Not in the next five years." Interoperability is not a nice-to-have.

He leaves listeners with artifacts rather than a slogan. A white paper spells out the three-layer stack. A clickable demo, built on the SRE example from the top of the conversation, shows an outage handled by agents that share cognition with counterparts in a cloud provider, a marketing team watching brand reputation, and finance watching cost. You can walk agent by agent through activity, intent, context, and the collective reasoning across the system. "A working demo is the coin of the realm," Labenz says. The twenty-agent SRE scoreboard is already production. The four-organization patient flow is already a demo. The missing layer is the one they are trying to write down.