Agentic IT Team update! · AI Automation Society

Agentic IT Team update!

A couple of months ago I gave one agent the keys to my entire home lab. Edge, mesh, LAN, compute, security, vuln scanning. Full SSH into my Proxmox cluster and workstations, the UniFi API key, Cloudflare, NetBird. Everything.

It worked. But it carried over 400 lines of code it had to read every run just to know what it could do, and one agent holding every credential is exactly the blast radius you do not want.

So I split it into 11 agents: least privilege, two jobs each, only the access those two jobs needed. Tighter, faster, safer.

Then the real problem showed up. An agent would get stuck in a loop, or keep calling a tool that failed, and I had no visibility into it. No run-level view, no cost number. I found out when something downstream felt off, not from anything I could open and read.

Observability was the missing piece of the puzzle. So I rebuilt the fleet on LangGraph + LangSmith. What that bought:

- Every run is a trace: which specialist ran, every tool call, tokens, cost. The stuck loop and the failing tool are visible in real time instead of after the damage.

- One supervisor is the only entry point and routes to the least-privilege specialist. "The supervisor is the only door" is structural, not a line in a prompt.

- Every change is proposed by one model and reviewed by a model from a different vendor (no model marks its own homework), then it pauses for a human.

- Everything writes to an append-only audit trail the agents cannot edit.

- I wired it into Slack, so I invoke any agent, watch it work, and approve a change with a button, right from my phone. I can run the whole home lab from anywhere.

The lesson, if you are pointing agents at real infrastructure: one agent with every key teaches you fast, least privilege makes it safer, but observability and cost tracking are what make it something you can actually run. Build those in on day one, and make the safety rails part of the graph, not the prompt.

Runbook

10 comments

Agentic IT Team update!