Use case · Operations

The night shift
that sleeps when you can.

An ops team deploys one hosted agent against the alert stream. It wakes when something happens, does the first twenty minutes of every response, and bills only for the moments it was awake.

An on-call rotation pays a human to sleep badly next to a laptop. Most nights nothing happens the cost is the vigil, not the work. A hosted agent inverts that. It costs nothing until something happens, then does the first twenty minutes of the response instantly.

1:13 AM

The pager fires. Nobody wakes up.

A p99 latency alert lands on the checkout service. The hosted agent — deployed once, subscribed to the alert stream, asleep since 11 PM — wakes on the message. No polling loop, no idle server humming all night for this moment.

It pulls the graphs, the recent deploys, and the runbook. The SRE on call is still asleep, which tonight is the correct state for an SRE.

a persistent agent's afternoon

idle — costing nothing

→ “what changed in our competitors' pricing this week?”

1:16 AM

The first twenty minutes, done in three.

The agent does the part of incident response that is reading, not judgment: it correlates the alert with a cache-node restart four minutes earlier, matches the runbook, and drafts the incident timeline while the incident is still young.

The one risky action — restarting the cache tier — is an external write, so it's held for approval, not fired into production at 1 AM. The metrics recover on their own by 1:31. Nobody was paged.

incident draft · inc-2214

01:13p99 latency alert — checkout service
01:14correlated: cache-node restart at 01:09
01:15runbook matched: cache stampede, § 4
01:16timeline drafted · rollback held for approval

6:00 AM

The morning report files itself.

A schedule fires at six, every day, whether the night was quiet or not. The report is already structured: 31 alerts overnight, 29 correlated to known noise and annotated, 2 promoted to incidents — each with a drafted timeline and the held actions listed.

The team reads it with coffee instead of reconstructing the night from dashboards.

schedule: "0 6 * * *" — the morning report

Mon 06:00

$0.14

Tue 06:00

$0.14

Wed 06:00

$0.14

Thu 06:00

$0.14

9:02 AM

The bill for a night of cover.

The ledger tells the whole night in four kinds of rows: the agent woke five times, ran about two minutes of metered compute in total, and the eight idle hours in between cost exactly nothing.

Every triage note, held action, and the 6 AM report itself is on the append-only trail. When the postmortem asks what was known at 1:16, the answer is a receipt, not a recollection.

the ledger never lies — append-only

holdreserved before the run−$3.00
charge43 GPU-seconds metered−$0.86
releaseunused hold returned+$2.14
receiptsigned, immutable, yours·
An outsourced overnight NOC starts around $3,000 a month. Last night's cover triaged 31 alerts, documented two incidents, and filed the report at 6:00 sharp — for $2.88. The hours where nothing happened cost what nothing should.
The uncomfortable math

0

alerts triaged overnight

0

real incidents, timelines drafted

$0.00

the whole night, on the ledger

The same shape, everywhere

Anywhere someone is paid to watch and wait, a hosted agent wins.

Queue and backlog watch

Support, review, and approval queues monitored; the aging ones escalated with context.

Vendor status tracking

Third-party status pages and webhooks watched so your customers never tell you first.

Nightly reconciliation

Payments matched against the ledger each night; only the mismatches reach a human.

Certificate and quota patrol

Expiring certs, filling disks, and rate-limit ceilings flagged weeks early, on schedule.

Release-night cover

An agent watches error budgets after each deploy and drafts the rollback request itself.

Daily standup brief

Tickets, merges, and alerts from the last 24 hours condensed before the team sits down.

Keep exploring

Give your agent a workforce.

Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.