Oluwatosin Oladayo Aramide

Part 1 of the Agentic NOC series. It introduces the two protocols that let AI agents reach tools and each other, MCP (the vertical axis) and A2A (the horizontal axis), and the third, ACP, that converged into A2A, plus the mental model the whole series is built on.

One HTTPS request followed from an external client to a pod, with the show commands and packet captures from every device it touches: BGP unnumbered underlay, type-2, type-3 and type-5 EVPN routes, symmetric routing proven on the wire, a Palo Alto A/A pair and Cilium.

There is a network inside a GPU server that moves more data than your entire fabric, and you will never configure it. Where NVLink ends is where your job begins: the scale-up versus scale-out boundary, and how GB200 NVL72 moved it.

At a few thousand GPUs, the backend network becomes a probability problem: something is always broken and the hash is always unlucky. The fix is two identical fabrics with zero cross-links. A deep dive into dual-plane design, sized for B300 nodes with 800G per GPU.

Two elephant flows, six trials, six collisions: one uplink carried everything while its twin sat at zero, and every dashboard stayed green. The measured polarization, the four-witness diagnostic chase, and the birthday math production fabrics live with.

Your AI cluster is two networks wearing one badge. One is judged by whether users notice it. The other is judged by whether GPUs wait on it. Frontend vs backend: the traffic, the rules, and how to design the GPU fabric so compute never waits.

5,412 connections on my Kubernetes cluster versus one RDMA flow moving 59.5 GB a minute. Then three identical transfers, one degraded path, and 32 percent of the workers' time gone. AI traffic measured, not described.

Credit-based flow control, a daemon that runs the whole fabric, adapters that hand out pointers into remote RAM, and switches that do math. Every InfiniBand idea an Ethernet engineer should know, demonstrated live on a real InfiniBand fabric.

The finale: one request traced live through all eight hops the series built, the five traps ranked by hours lost, and what I would do differently next time. Plus the honest single point of failure the design still carries.