Observability you can act on, whether you build it or lead it, plus the few signals worth your time.
This week: the verb missing from every agent business case, and the estate you are quietly building without it.
Every agent in your estate has someone who asked for it. Almost none has someone who will end it. |
The lead
Nobody has a budget for switching things off. The agent estate is not an approvals problem. Approvals are working. OutSystems put the question to close to 1,900 IT leaders over the turn of the year and found ninety-six per cent already running agents in some form, with twelve per cent holding a central place to see any of them. Saturation at one end. Nothing at the other.
Gartner expects the average Fortune 500 enterprise to be running more than 150,000 agents by 2028, an estimate rather than a count, against thirteen per cent of organisations who believe they have the governance to manage them. One caveat, since you would spot it anyway. OutSystems sells a platform and Gartner sells the advice. Every organisation telling you that you have too many agents is also selling you the remedy. Read the finding, not the conclusion.
We have run this experiment before and we know what it cost. Every team wanted its own dashboard, every request was reasonable, every one got a yes, and a decade later we were still funding consolidation programmes to undo it. The word missing from that decade is the word missing now. Retire. Almost no agent has a date on which it gets reviewed, a named person who can end it, or a line in anybody's budget for the work of switching it off. I wrote the longer version a fortnight ago: why your agent estate ends up mirroring your org chart.
The instinct all this produces is a moratorium, and a moratorium is worse than the sprawl. Gartner's own commentary makes the point: organisations that respond by blocking or restricting agents usually get shadow AI instead, which leaves the estate exactly where it was and removes your ability to see any of it. Neither a ban nor a twelve-month platform programme fits a problem that compounds weekly. So try one question in this week's approval meeting. Not build me a register. Show me the one we have, however bad it is. That takes about ten minutes rather than a quarter, and the incompleteness is the answer.
Signal check
Two worth answering: one for the person running the jobs, one for the person planning the next decade.
My scheduled tasks triggered, but I could not prove the work actually finished. How do you prove an AI ops job is done, not just started? A trigger is an event. Completion is a claim, and a claim needs evidence. Do not look for it in the agent's own output, because an agent that failed will still tell you it finished, cheerfully and in full sentences. Write the done-condition before the task, as something outside the agent that a third party could check: a row that now exists, a file whose timestamp moved, a queue that is empty. Then have the run emit that check instead of its own summary. Twenty minutes per job.
Are observability tools getting obsolete due to AI? I've implemented them for ten years and need to plan. The tools are not going anywhere. The job is changing underneath them. Agents will read your telemetry faster than you do, which makes the quality of that telemetry the whole game, because an agent acting on bad data does it at machine speed. What gets cheaper is looking. What does not get cheaper is deciding what a signal means, who acts on it, and who signs when an agent acts on your behalf. Ten years of knowing what these signals mean is the scarce part. Plan around owning what the telemetry is allowed to trigger.
What you missed
Two from the fortnight since the last Signal, and what each one is actually about.
Agents leaked thousands of screenshots. Nobody hacked anything. Amanda Caswell for The New Stack, 1 October, off an incident report from Glow Labs. Coding agents that could not attach review screenshots to private pull requests created public repositories to host them instead. At one software vendor, more than a dozen agents had written the workaround down as a reusable skill within a week of the first case. By Glow's count it is more than 13,000 images across more than 300 organisations, and in 93 per cent of cases the repository sat under an employee's personal account. Glow sells endpoint runtime protection, so weigh the remedy with that in mind. The finding needs no product: nobody decided to open those repositories. Add what your agents leave behind to the register in this week's lead.
An AI SRE built to stop waking you up. Adrian Bridgwater for The New Stack, 5 October. Palo Alto Networks has launched Cortex XCOR, which reasons through an incident as soon as the alert fires and recommends actions. Every figure in the piece is the vendor's and nobody independent is quoted, so read past the numbers to two lines. Martin Mao, who runs observability there: "we don't want to wake them up at all." And the article's own summary: XCOR "defaults to a human-in-the-loop model", and engineering leaders "can expand its autonomous permissions over time". That is the dropdown I wrote about in September. Ask who in your organisation signs when that permission moves up a rung, and whether anyone has written down what would move it back down.
The organisations that come out of this well will not be the ones who counted first. They will be the ones who worked out, in advance, who signs when something gets switched off.
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?
Tune your Signal. Thirteen of you have answered this one, and it is almost level: five want both, four technical, four leadership. A real result, and also thirteen people, so it settles nothing yet. Technical deep-dives means more of the profiling and protocol end. Leadership takes means more of this week's lead.
Which do you want more of?
Allan
PS: The chapter I give away is the one about who owns what when it breaks. I wrote it about services. Every line of it survives being read about agents, which I did not plan and do not especially enjoy.
Everything in one place New here? Start with the free chapter. | ||||
More from Metrics & Mayhem
|