Analyzing UniFi with an AI agent: my Hermes experience in the home network

I have a UniFi setup that I recently rolled out across the whole house: a few access points, one switch too many, VLANs I once wanted to set up "properly" and then mostly slapped together.

Works.

But "works" is not the same as "good".

That was exactly the part that annoyed me. The WiFi was there. The devices were online. The UniFi interface did not look like anything was currently on fire.

Still, I had this feeling: something is probably wrong somewhere, I just do not see it right now.

So I did something that felt absurd at first: I let an AI agent, specifically Hermes, loose on my network and asked it to check it.

I did not care whether that sounded like the future.

I cared whether anything would actually be better afterwards.

Why Hermes was interesting in the first place

I had not completely written off AI agents before, but I was not really convinced either.

My first attempts with agents often felt like demos: exciting, a bit wobbly, a lot of fiddling, and in the end I was sitting there again doing the actual work myself.

Hermes was the first tool in that direction where it felt different.

Not perfect.

Not magic.

But calm enough to give it a real task.

And a UniFi home network is a pretty good test for that. There is real data, real configuration and real consequences. If an agent only throws buzzwords around there, you notice quickly.

What I did

Not much, actually.

I created a local read-only user in my UniFi console and gave the agent the credentials plus the IP of my Dream Router.

From there, it logged in independently, got a session and checked out the necessary APIs.

What the agent actually did

Hermes pulled an overview through the UniFi API.

Devices.

Firmware versions.

Channels.

Transmit power.

VLAN assignments.

Firewall rules.

That sounds dry at first. It was.

But that was exactly the good part.

The agent did not just unpack the classic reflex and recommend "more power" everywhere. It looked for patterns: which access points were interfering with each other? Which devices were connected to the wrong AP? Which settings looked like leftovers?

That was the moment where "AI plaything" became a useful audit.

The things that actually made a difference

• Untangled the channels. Auto is not always smart. Fixed, non-overlapping channels noticeably lowered latency in the living room.
• Transmit power down, not up. Sounds wrong, but it is right: less power, cleaner cell boundaries, better roaming transitions.
• Pinned devices to one AP. Some clients stubbornly kept connecting to the much farther away access point.
• Made leftovers more visible. Not everything was critical. But a few settings were only still there because I had created them at some point and then forgotten them.

The most important point was not a single magical recommendation.

The most important point was that someone, or something, stubbornly checked every corner.

What you should know

I would not let an agent blindly make changes to the network.

Absolutely not.

A home network is not a text document where you briefly press undo and everything is fine again. If an agent misunderstands a firewall rule, misclassifies a VLAN or makes an assumption about devices, that can get annoying.

For me, Hermes works here as a very thorough second look.

It can check, sort, suggest and explain.

I still have to decide.

That is not a downside. That is exactly how I want it.

Would I do it again?

Yes, and regularly.

Not because I could not do it myself.

Because an agent stubbornly checks every corner while I say "good enough" after ten minutes.

That stubbornness is the value.

What I learned: the agent does not replace understanding. It replaces gut feeling with data.

And my gut feeling was wrong in two out of three cases.