An agent that administers the network
what it can do, and its limits

What an agent is, concretely

A language model alone answers questions. An agent additionally has tools, execute a command, read a file, query an API, open a page, a persistent memory, and the ability to chain these calls to reach a goal. It is this loop that produces a verifiable result: a report, a fix, a status overview.

The workstation, and the scope

We run this agent on a Linux admin workstation, in a private meshed network connecting servers, the admin workstation and client sites. Nothing is exposed to the Internet: the agent only reaches a machine if the mesh allows it.

The guardrails we apply

• Read before write the agent observes, measures and reports. Any action that modifies a system goes through an explicit human decision.
• One action at a time no script touching ten machines at once without having verified the first one.
• Post-action verification the report of an action is not proof of its effect: we re-read the actual state, we do not trust the report.
• No secrets in the context credentials stay in protected files on the workstation, never in the conversation.

What it really changes

The gain is not in decision-making, it is in collection: surveying the state of fifteen machines, cross-referencing logs, producing a dated and encrypted report. This work took half a day; it takes a few minutes, and the administrator spends the remaining time on analysis and arbitration.

The downside is symmetric: an agent that makes a mistake makes it quickly and with confidence. That is why every intervention ends with a live verification, and with a written report the client can review.