Why it matters
For whoever picks the tools
Your telemetry has a new reader. It doesn't need the whole picture up front.
At 2:14am the dashboard renders perfectly — traces, cardinality, the graph showing which service turned red and when. It was all gathered in advance, because the person it was drawn for arrives cold and cannot go and look.
The shift
For whoever picks the tools
Breadth up front is what you buy when the reader arrives late.
A person cannot ask the machine anything, so everything has to be gathered up front. The new reader was already on the box when the event fired — it fetches exactly what this event turns on: which process holds the file, whether the service answers right now. Some of that was never in your pipeline. So the shape inverts. Instead of collecting everything in case someone needs it, the agent keeps a narrow set of core signals running and widens only when one moves — on the host, where the data already is, so what crosses the network is the finding, not the window it came from.
Monitoring
Is it alive?
A threshold and a page — built for a person on call, and honest about it: a tripwire.
Observability
Why did it break?
Twenty years spent explaining an unfamiliar failure to a person. Gathered wide and kept, because the person would arrive after the fact.
AgenticOps
What does this one need — and did it work?
A reader already on the host asks a narrower question, answers it, then acts. What counts is no longer how much was collected. It is what was fetched.
A person needs it gathered in advance, because by the time they read it they cannot go and look.
The agent was already there. It can just look.
Explaining still matters. It just moves: your engineer does the explaining once, in daylight, by naming the commands — which process must never be killed, which host is not touched on a Tuesday. We are not the only ones saying this. The category has said versions of it for a year, and the shift keeps accelerating: most of the industry is still watching; fewer are running. Analysts tracking this space describe the same endpoint — manual incident work fading, not gone, with less of what's left done by a person's hands. Two questions are still open about who delivers it: quickly (how many layers between the alert and the host?) and safely (is the risk known before it runs, not after?).
Question one · Fast
For whoever counts the moving parts
How many layers sit between the alert and the host?
Count the processes a signal has to cross before it changes anything on a machine — each one is a place to be misconfigured, or to wait.
The usual shape
The observer and the actor are different systems, so a tier has to sit between them to carry intent across: a connection, a runner on the target, a second set of credentials, an allowlist.
Here
The agent that noticed is the agent that acts — one service, on the same machine. That tier is not simplified. It never gets built. Nothing has to be kept in sync with reality, because it is reality.
This is architecture, not a feature. It stays true when someone else ships an integration. Count the artifacts for one ordinary fix →
Question two · Trust, before it runs
For the engineer who has to approve it
Is the risk knowable at review time?
"Is it safe?" usually gets answered after the fact. The better question: can the person reading an action tell what it is capable of while reading it, before anyone arms it?
Constrained language
No shell escape. An action calls only Builtin Functions from its registered catalog, so the list of Builtin Functions it calls is the exact extent of what it can touch.
Graded by reversibility
Each Builtin Function carries a static risk grade — whether the thing can be undone, not how dramatic it sounds.
The gate reads the grade
That grade and who is asking settle whether something may run unattended — not how urgent the alert looked.
Armed deliberately
Library actions ship disarmed. Someone turns one on, separately from installing anything.
The third question
For both of them
And did it work? Nobody has answered this one.
There is a third question everyone skips: did it actually work? Not "did the command exit zero" — did the incident resolve, against a criterion agreed in advance.