Watch
The incident opens itself.
Watch is health, errors and incidents, correlated with releases: monitors, error rates and deploys on one timeline, so when a service degrades, what breached is named against the release it moved with.
An incident here is not a notification to acknowledge — it is where the autonomous loop begins.
4m 13sfrom the first bad signal to the merged fix — the run the landing replays, measured end to end.
Health, errors and incidents.
Each signal keeps its own figures, and the incident timeline is where they meet — with every deploy on it, so nothing is read in isolation.
Monitors
Probes against the endpoints you ask Datis to watch. A failure is recorded as what the probe got back and how many times in a row — including the service that answers on time and writes nothing.
Error rates
Errors arrive grouped, in the service’s own words — the type, the message, and the tally of events and users behind it. A service’s rate is measured against its own thresholds.
Releases
Every deploy lands on the same timeline — the pull request it came from and how far it rolled out — so a degraded service is read against the release it moved with.
Failures have to agree.
Datis does not open an incident on one bad probe. Two things have to be true at the same time — and until both are, nothing moves.
More than one monitor is failing
A single failing probe opens nothing. Watch waits until another monitor is failing too.
The error rate has breached
Failing probes alone are still not an incident. The service’s own error rate has to be over its threshold at the same time.
The incident opens itself
Both at once, and the incident opens on its own: it names what breached, pages whoever is on call, and the clock starts.
The timeline is where the loop begins.
What Watch writes down is what the agent picks up. Triage does not reason from the alert alone — it starts from the record the incident already holds.
Watch one incident run end to end on the landing- The clock
- It starts the moment the incident opens. The run the landing replays is measured from here — the first bad signal — to the merged fix.
- What the agent reads
- The errors, the recent releases, the code in the path, your runbook — what a person would read, before it names the change that did it.
- What comes back
- Agent runs land on the same timeline with their outcome named: merged under a policy that earned it, or waiting for a person — and the ledger records which of the two happened.
Connect what you run. Put it under watch.
The beta is private and the waitlist is the door: joining creates no account and costs nothing, a person reads every request, and design partners step in first.