The Internet of Agents (This is What's Coming)

Most enterprises think they're further along on AI than they are. Cisco opened its AI4 2026 session with a blunt map of where companies sit — and the hard part, according to their framework, is still ahead.

Where Most Enterprises Are

Cisco laid out four stages. Most companies today are somewhere in the middle two:

• Experimentation — pilots, isolated use cases, vendor-led deployments through OpenAI, Microsoft, or Google, with limited governance.
• Initial Production — early multi-agent orchestration starts, governance and compliance programs begin to emerge, and the tooling shifts toward LangGraph, AutoGen, Semantic Kernel, and vector databases.
• Multi-System Orchestration — AI spreads across business units, multi-cloud and API sprawl sets in, and agent sprawl plus identity fragmentation becomes a real operational headache. Tooling here looks like Kubernetes, Kafka, API gateways, and Datadog.
• Cross-Enterprise Federation — the stage nobody's really at yet: agents interacting directly with partner organizations' agents, needing regulatory-grade auditability and solving distributed trust problems that don't have established answers.

The framing matters because most AI vendor pitches sell as if federation is around the corner. Cisco's own map puts most companies one or two stages before that, still fighting agent sprawl and identity fragmentation inside their own walls.

Three Transitions Happening in Enterprise AI at Once

Cisco broke the shift into three separate axes, each one a real engineering and organizational problem on its own:

Topology — centralized to decentralized. Instead of one agent (or one supervisor agent) doing everything, specialists each own a piece of the problem. A single agent is simple, fast, and cheap — good for document summarization or support Q&A. A supervisor pattern decomposes complex work like RFP responses or financial analysis into pieces with centralized oversight. A peer-to-peer pattern, where agents collaborate directly with bidirectional links, is built for things like collaborative code debugging or market analysis — highly scalable, but harder to govern.

Execution — static to dynamic. Static workflows follow predefined steps and break the moment something unexpected happens: step one, step two, step three fails, stop. Dynamic workflows assemble the right team for the specific situation and adapt as it evolves — understanding the goal, discovering the right agents and tools, planning an approach, and adjusting in real time. The gap with static workflows is real: rigid paths can't handle novel situations, failures require manual rework, and none of it scales cleanly across different contexts.

Agency — triggered to autonomous. Triggered agents wait for a person to ask, then respond. Autonomous agents detect signals, decide among options, and act with actual authority, in a continuous loop. The tradeoff with staying triggered-only: every action needs a human to start it, there's no authority to commit to anything, latency costs you opportunities, and the system never learns from its own outcomes.

The Real Problem: Agents Can Connect, But They Can't Coordinate

Here's the part of the session that's easy to miss under all the framework language, and it's the most important point in the whole talk. Multi-agent systems today come from different vendors, run on different frameworks, live on different clouds, and operate under different access controls. They have no shared memory, no shared context, no shared goals, and no way to negotiate with each other.

Cisco's own words on it: agents "can connect together" but "cannot think together." Message passing between agents isn't collaboration — it's just "putting stuff out the window and assuming that person understood it and acts on it."

Why the Internet of Agents Framework Exists

Cisco's proposed answer is what they're calling an "Internet of Agents" — an open, interoperable layer for connecting multi-agent systems, built around three needs: the ability to discover and identify other agents and judge whether they're reputable enough to act on your behalf; the ability to securely connect and exchange text, video, image, or audio state across systems; and the ability to observe and evaluate whether agents are actually doing what they're supposed to do, or getting stuck in loops and conflicts.

None of this is solved yet industry-wide. It's a real, open problem — which is exactly why it's worth watching.

What This Means Right Now

If you're evaluating multi-agent AI tools for your business, don't assume interoperability between vendors exists yet, because it mostly doesn't. Know which of the three transitions (topology, execution, agency) your use case actually needs before you buy anything — a lot of vendors are selling autonomous, decentralized systems for problems that only need a single, triggered agent. Simpler is usually cheaper and easier to govern.

Trying to figure out where your business actually sits on theAI maturity curve? Winsome helps companies cut through the framework noise and figure out what stage they're really in — and what to build next. Talk to Winsome about your AI roadmap.

 Source: Cisco's Ai4 2026 session, "Internet of Agents"