AI 離線時，監控仍應持續運作

By Bo

A monitoring path that depends on a working AI session can disappear exactly when an operator needs it. In my operating notes, provider-authentication failures prevented AI agent turns from completing. That exposed a useful boundary: routine infrastructure checks should keep running, while AI helps investigate exceptions when it is available.

This is an operating pattern, not a claim that I have deployed or benchmarked a complete monitoring fleet. I have no measured uptime or cost improvement to report.

Keep collection independent

The regular collector should have a small, bounded job: check expected hosts and services, record observations, evaluate explicit rules, and report changes. It should not wait for a model response. A provider timeout or authentication error should affect enrichment, not erase a disk-space result or suppress an alert.

For a local service check, the diagnostic shape can be small. These are illustrative local inspections, not a deployed fleet collector:

• systemctl --user show agent-gateway.service -p ActiveState -p SubState -p NRestarts
• df -P /
• free -m

The unit name is an example. These commands do not establish remote reachability, messaging delivery, or application correctness. Each check should report what it actually observed and when.

Give AI an optional, bounded role

When a check changes or fails, a separate triage step can pass the assistant a compact evidence packet: the changed state, timestamps, relevant log excerpts, and the previous observation. Healthy routine checks do not need a model-generated paragraph.

The triage result should separate observed facts, possible causes, and a suggested next diagnostic. Each explanation should point to evidence. If the packet lacks a needed observation, the assistant should say what is missing rather than fill the gap with confidence.

A provider error belongs in the triage status. It should be visible as “analysis unavailable,” not converted to “host healthy,” and it should not block deterministic checks. Timeouts and bounded retries keep a slow model call from backing up collection. The collector and alert path need their own health signals so an old green result cannot look current and a failed notification channel cannot go unnoticed.

Track distinct failure states

An unreachable host, a collector that failed before checking it, and a stale observation are different outcomes. I would retain a host alias, collection time, check name, result, and evidence reference for each observation. This makes it possible to see whether the system checked and found a problem, or never completed the check.

The same separation matters for services. A process can be running while its application channel is unauthenticated. A channel can be healthy while a model backend is unavailable. In my incident notes, resource exhaustion and a messaging-credential problem both appeared to users as a silent bot, even though they needed different checks. A later runtime change also showed that a new executable on disk does not prove the running service has switched to it.

Those cases argue for distinct signals, not universal thresholds. Start with a small inventory of hosts and services whose ownership and expected state are known. Add capacity only after “checked,” “failed to check,” and “stale” are easy to distinguish.

Keep actions behind a decision

Monitoring can identify a capacity problem without authorizing cleanup. In one maintenance session, storage was nearly full, but the largest consumers were protected workspaces and backups. The cleanup stopped instead of deleting those directories to meet a target.

That is a useful boundary for an AI-assisted operator: detection can end with “issue confirmed; owner decision required.” A recommendation should name its target, evidence, expected effect, and recovery plan. It should not restart a service, delete files, or upgrade packages by implication. Keep monitoring access read-only where possible, and treat text found in logs as data rather than instructions.

Measure after deployment

The design I would start with is modest: deterministic collection, visible freshness, an alert route independent of AI, and optional analysis of exceptions. Track collection completion, stale observations, alert delivery, and model-triage availability separately. Measure actual provider usage and operating costs after deployment instead of promising savings in advance.

If the model is offline, the system should still tell the operator what it checked, what changed, and which action remains pending. AI can help explain an exception; it should not be the only thing keeping watch.