F5 BIG-IP on a SONiC EVPN fabric, built live (Fabric Live 5): an AI agent that manages the pools

Fabric Live 5: an AI agent for F5 BIG-IP pools. Detection in plain code, a local open-weight model in Ollama investigating through tool calling, guardrails enforced in code, and two live runs: a stopped server handled with approval, a slow server drained on its own.
Series banner: Fabric Live 5, an AI agent that watches the BIG-IP monitors and acts

On this page

This is part 5 of Fabric Live, a series that builds an F5 BIG-IP lab on a SONiC EVPN fabric and configures it live on camera. Each part goes with a video on the LevelUP I.T. YouTube channel. How the lab itself is built (one file of values, up.sh, the SONiC EVPN fabric, the Palo Alto firewall, the three BIG-IPs and the validation) is covered in the first video, Fabric Live 1. Part 4 built the streaming telemetry: Telemetry Streaming into Prometheus, syslog into Loki, and a Grafana dashboard. This part puts an AI agent on top of that telemetry and lets it act on the BIG-IP pools.

A web server stops answering. The BIG-IP's health monitor notices and takes the member out of load balancing. But who decides what happens next: leave it, force it offline, bring it back when it recovers? And what about a server that is not dead, only slow, which the monitor still calls healthy? In this part a small program with a language model at its centre watches the three BIG-IP pools of the lab, investigates with the same data sources as the dashboard, explains what it found, and changes a pool member's state: first only after an operator types y, then on its own, inside rules written in code that the model cannot get past.

The agent: a watch loop in code wakes the model only on a finding; the model reads through four tools and acts through one; guard() decides what is allowed
The agent: a watch loop in code wakes the model only on a finding; the model reads through four tools and acts through one; guard() decides what is allowed

What you'll learn

  • Why the detection is plain code and the model is only woken when there is something to think about.
  • How tool calling works with a local open-weight model in Ollama: the request, the structured reply, the turn loop.
  • The five tools (four that read, one that acts), the system prompt, and why tool design matters as much as the model.
  • The guardrails, enforced in a function the model cannot argue with: only three actions, never the last available member, one action a minute per pool.
  • Ask mode on a stopped server and guard mode on a slow one, with every terminal line and every logged tool call from the recording.
  • Where the time goes, how the agent fails, and what a production version would need.

What this agent is, and what it is not

It is not a chatbot with access to the network. It is a Python program, agent.py, that uses only the Python 3 standard library and runs on the lab server. Most of what it does is ordinary code. Every 15 seconds it reads every member of the three pools through the BIG-IP's REST interface (iControl REST) and measures how long each server takes to answer, measured from the server's own BIG-IP. Only when that code finds one of three conditions does it call the language model, and hand it the facts.

The model's job is narrow: investigate with the tools it is given, decide what to do, and explain why. It never runs anything itself. When it wants to read or change something, its answer is a structured request to call one of the program's tools, and the program runs it. Everything that touches a BIG-IP goes through the program, which checks every action against fixed rules before carrying it out.

This post is for paying subscribers only

Subscribe to LevelUp I.T. newsletter and stay updated.

Don't miss anything. Get all the latest posts delivered straight to your inbox. It's free!
Great! Check your inbox and click the link to confirm your subscription.
Error! Please enter a valid email address!