Observability you can act on, whether you build it or lead it, plus the few signals worth your time.
Back after a fortnight: the accuracy figure in the demo was graded by the vendor who made it, so here is the test you can run instead.
|
You can automate the action. You cannot automate the accountability. |
The lead
Which rung is this on, really? That is the question I asked a vendor this week, and I did not leave without the answer: read-only, the bottom rung. AI-SRE is two words stretched across four different things. At the bottom, something reads your telemetry and writes you a tidy summary. At the top, something restarts a database on its own while nobody is awake to stop it. If you cannot tell which the demo is showing you, you cannot know what you have signed up to.
The ambition is right. The first half hour of an incident is mostly gathering before one moment of judgement, and almost none of it is clever work. A machine does it better than we do, because it does not get worse at four in the morning. So what needs settling is not whether to do this. It is where on the ladder you sit, and that is a governance decision rather than a purchase.
So test the working, not the accuracy. A hypothesis should arrive with its evidence attached: this deploy, this error rate, this dependency, the past incident that looked like it. Hand me a conclusion with nothing behind it and I am taking a stranger's word for it on the worst night of the quarter. Same for the guardrails. There are three. Bounded, so it restarts one service and not the cluster. Reversible, so a rollback is fair game and a deletion is not. Approved, so anything with real consequence waits for a person. The OWASP Top 10 for LLM Applications names excessive agency a risk in its own right, and a human approving high-impact actions as the mitigation. If a vendor cannot walk you through all three, they have built a confident chatbot holding production credentials.
I used to tell people to automate everything they could, and I learned that rule's failure mode the expensive way. We automated something, we did not measure the blast radius properly, and it cost us the client. You can automate the action. You cannot automate the accountability. So the evaluation I would run needs no procurement: one read-only agent, pointed at your estate, writing the incident timeline for a month with the authority to touch nothing. Then set its timeline beside yours and keep two numbers. How often it found the cause before you did, and how often it was wrong while sounding certain. The four rungs, and where each breaks.
Or listen on Tech Tuesday 5:
METRICS & MAYHEM
What Is an AI-SRE? The Autonomous Ops Reality
Signal check
One from the floor, one from the review meeting.
How do you tell when an agent has stopped behaving like itself? Not by reading its output, which is where most teams start. Watch the shape of what it does. Record, per run, the tools it called, in what order, how often, and how long it took to reach a conclusion. That is your baseline. Drift shows in the shape before it shows in a wrong answer: the same question taking nine tool calls where it took three, or reaching for something it has not touched in weeks.
How do I spot that one of our KPIs has become a target being gamed, before somebody else points it out? Goodhart's Law gets quoted here, usually too late to help. The practical version: an unpaired number is being gamed somewhere, and you can go and find it. Take your top ops KPI and name the laziest way to move it without doing the job. Reduce MTTR, unpaired, rewards whoever closes the ticket fastest. Cut alert volume, unpaired, rewards silencing alerts. Add the measure that catches that move, and add it before the target goes live. Which is the lead in miniature: a vendor's accuracy figure is an unpaired number, and the pair it lacks is how often it was confidently wrong.
Latest on the blog
One, and new since you last heard from me. If a renewal is what is on your desk instead, what to do when your vendor buys your other vendor is about the bargaining power that moves quietly across the table on consolidation, and what to ask for while you still hold some.
What you missed
Three worth catching up on.
Most teams already trust an agent they cannot grade. Splunk's survey of 2,050 IT and engineering professionals across ten countries, reported 22 October by Anamarija Pogorelec for Help Net Security. Sixty-two per cent regularly rely on AI agents to work out why their systems broke, and fifty-six per cent cannot measure whether those agents get it right. Thirty-eight per cent have not decided who owns agent observability, which Splunk's Mark Woods puts plainly: "When agent observability is everyone's job, it becomes nobody's job." Splunk's own survey figures, not an independent measurement. The survey's findings on invented answers.
Grafana's case for an error budget on agent behaviour. Craig Risi for InfoQ, 16 October: Grafana turns each behavioural check (grounding, fulfilment, toxicity, personal data exposure, prompt injection) into a separate SLI. The worked example is 95 per cent of conversations judged fulfilled over 30 days, a 5 per cent budget. As Risi puts it, "An SLO is only as useful as the measurement underneath it." Who owns that budget? The scoring behind the budget.
Somebody has open-sourced the watching part. James Maguire for Security Boulevard, 28 October, on Cisco's Project Quarry, which judges a whole agent run rather than single commands, built to catch an attack made of steps that are each legitimate alone: read the credentials file, then post to an unfamiliar host. Cisco calls its effectiveness unproven. Rare honesty. Quarry's whole-run rules and sample plugins.
Before you go
Reply and tell me the observability problem on your desk this week. I read every reply. The best ones become a Signal check, and if it is a bigger piece of work, that is exactly the kind of thing I help teams with.
How useful was today's Signal?
Allan
PS. New and free: The Outcome Question, a twenty-minute test for the number your CFO watches, and The Signal Audit, thirty minutes on your own last incident.
Before the next demo, take the incident that actually hurt you this year, the novel one nobody had an alert for, and write down what an agent would have needed to reach your conclusion. Whatever is missing from that list is what no demo will fix.
|
Everything in one place New here? Start with the free chapter. | ||||
More from Metrics & Mayhem
|