Agents and guardrails
Picks checked October 2026.An agent is a model that decides which tool to run next, runs it, reads the result, and keeps going. On a network that's a big step up from a chat window, so the way I'd start is read-only first, tools that do the same thing every time, and a human who approves anything that changes the network.
Where I'd go
Building effective agents, Anthropic
Where to start, and when you shouldn't build an agent at all. I'd read this before writing a line of code.
Writing effective tools for agents, Anthropic
How to build a few good tools that return the whole answer, so the model picks the tool and the tool does the work. It's the part most demos skip.
LLM06: Excessive Agency, OWASP Top 10 for LLM Applications
The security team's list: least privilege, read-only where you can, no open-ended tools like a shell, and human approval for anything high-impact. Handy to send your security reviewer.
Bounded network AI agent (video and write-up)
An agent that audits my lab, finds a real CVE against Cisco's live advisory feed, writes the fix with a rollback, and then can't apply it until a named human approves it. Every call it makes goes to an audit log.
It isn't finished. The approver's name comes through the model, so it enforces the process, not who you are. I say so in the write-up. It also needs a lab, NetBox, and Claude Code.
If you want to try it on a local model instead, my local agent harness post tests where the approval prompt actually lives (in one app, it didn’t).
I cover the network side there.
Left out on purpose
- Agent frameworks and the courses built on them. The Hugging Face Agents Course is free and good, but it teaches three frameworks before you've decided you need one.
- Vendor agent platforms. That's a buying decision, not a learning one, and this page isn't for selling you anything.
- Agent repos from social media. Most are posted by the person who wrote them, and I wouldn't point one at a network until I'd read every tool it exposes.
I'm still working out where approvals belong myself, so treat this as where I'd start today.Garrett