What an AI DevOps agent can fix alone
Every five minutes, Vigil looks at what its agents report: metrics into VictoriaMetrics, logs into Loki, each system carrying its own model of what normal looks like. When something drifts, the agent has two ways out. It can act, within a short list. For everything else it calls a human, and the call arrives with the numbers.
That split, what it may do alone and what comes back for review, is where the design work of an AI DevOps agent happens. Detection is comparison. Deciding what an agent may do without a person is a design decision, and design decisions deserve documents. This note is ours: where we draw the line for Vigil, and the reasoning behind it.
Vigil is a Kerber AI venture, live at vigil.kerber.ai. The product claims below come from there and from our ventures page, both read on November 9, 2026. No new experiment was run for this note. Where the note argues rather than describes, it states a design position, not a product promise.
A monitoring stack, minus the human in the middle
The machinery is deliberately ordinary. One docker-compose file brings up VictoriaMetrics, Loki and Grafana on an instance. A one-liner installs Node Exporter and Promtail on each server. From then on the stack reads itself: metrics as they stream, logs as they land, and Vigil on its own five-minute cycle above both.
Nothing about five minutes is a technical limit. It is an interval picked to catch a shape before it turns into an outage, and to make one bad data point harmless to a decision.
Thresholds page too late
A threshold pages you when a number crosses a line. That is usually hours after the number started moving, and it says nothing about whatever moves slowly. A memory leak does not page at 89%. It pages at 100%, often at night. A slow shift in error rate can hide behind a green dashboard while the total keeps sliding.
So Vigil's first question is not which line got crossed. It is what normal looks like on this system. The agent keeps a per-system model of ordinary behavior and reacts to drift from it: CPU trends, connection pool pressure, error rate shifts. What normal looks like is learned from history, which is the same argument we made about agent memory and work traces: history stored well is evidence, not clutter.
The line
Here is the whole list of actions Vigil takes without asking: restart services, scale capacity, clear queues. Three entries. The list is short on purpose, and every entry earns its place the same way: the cost of being wrong is bounded.
A restart puts a process back where it was. Scaling spends money, not data. For queue clears, the design principle is that an agent should only empty queues whose work is idempotent; where the work cannot simply run again, an emptied queue destroys anything half done. That principle governs the drafting of the list. Everything outside the list escalates, with full context.
We learned the hard way what happens when agents that can act have no boundary: our fallback chain once had all twelve agents trigger at once, overwhelm the local model and collapse into the cloud fallback. An on-call agent has to inherit that lesson at birth, because it acts while nobody is calm, or even awake.
One more design rule for any agent that can do things: the allowed list should live in code, not only in a prompt. A prompt helps the agent choose well. Enforcement is what stops the tool when it chooses badly. We wrote a full field note on that distinction while auditing an agent that sent a message no one had approved.
The rest comes back with context
An escalation carries what triggered it, the values around it and a root cause in plain language. The first screen you open can be a decision instead of a dashboard. Telegram hosts the thread. You ask follow-up questions right there, trigger actions from the chat, and silence the noise that turns out, on inspection, to be noise.
That changes who can respond. Not a specialist who must already be awake, but whoever holds the phone, with enough context to judge for themselves. The alert is an interface problem as much as an agent problem, and it is the part that decides whether the whole loop closes.
One stack, many systems
Vigil runs multi-tenant: one stack, full separation per product, and a new agent on a new server in under three minutes. Remediation is cloud-native for AWS and GCP, so the same agent and the same interface act on ECS, Lambda, Cloud Run and Compute Engine.
For a studio that runs several products in parallel, that ratio is the point. One operating model, portable to the next system in minutes.
What we keep for people
Some decisions stay human because they should. Whether to page tonight or wait for morning is a cost judgment, not a data point. Whether tonight's deploy is riskier than a restart at 2am is context no metric carries. Whether a queue should be emptied while its producer is down is a business question wearing a technical disguise.
And the permission list is only as good as the spec behind it: what the agent may decide, what normal means for each system, which drift to watch and which drift to stop. That writing is the job, as we say everywhere else in this studio. A good model with a bad spec makes a fast, confident mistake.
The loop itself is short to stand up: one docker-compose file, one one-liner per server, and an agent that reports to a chat. The published product shows it end to end, and the ventures page shows where Vigil sits next to the rest of what we build.
Want more? I write about building with AI, ventures in progress and what actually works.
No spam. Unsubscribe any time.