Automation
Set the schedule.
Skip the babysitting.
Cron schedules fire agents, swarms, and simulations with exactly-once semantics. Signed webhooks deliver the results, evaluations grade them, and artifacts keep them.
curl https://api.swarms.dev/api/v1/schedules \ -H "Authorization: Bearer $SWARMS_KEY" \ -d '{ "cron": "0 6 * * 1-5", "target": { "kind": "swarm", "swarmId": "swm_report" }, "budgetUsd": 0.75, "webhookUrl": "https://ops.example.com/hooks/report" }' # → 201 { "scheduleId": "sch_2ab…", "nextFire": "Mon 06:00 UTC" }# fires exactly once per tick — no doubles, no gapsThe most valuable work is the work that happens again — the 6am report, the nightly sweep, the weekly teardown. Swarms runs it on a schedule with exactly-once semantics, grades every output against a rubric, and delivers results over signed webhooks. Recurring work should not require a recurring human.
Schedules
Cron that fires exactly once. Every time.
A schedule can fire anything on the platform — a single agent, a 16-worker swarm, a full market simulation. Each tick runs with exactly-once semantics: a redeployed worker or a flaky network never produces a duplicate run, and never drops one.
Every tick is also a governed run. Each firing carries its own budget cap, so the report that runs 260 mornings a year costs what you decided — 260 times.
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
Webhooks & artifacts
Results that arrive signed and stay put.
When a run finishes, your systems hear about it on your terms: outbound webhooks are signed so you can verify the sender, and retried with backoff until acknowledged. A blip on your side delays delivery; it never loses it.
The output itself becomes an artifact — a durable result file with a retention policy, not a payload that vanishes with the response. Thursday’s report is still there when the question arrives in June.
POST https://ops.example.com/hooks/reportX-Swarms-Signature: t=1768543200,v1=9f2c8a… { "event": "run.completed", "runId": "run_7fe…", "costUsd": 0.61, "artifacts": ["art_5kq…"]} # retried with backoff until your endpoint acks# artifacts stay downloadable for their retention windowEvaluations & replay
Output that gets graded, then debugged.
Unattended work needs an unattended reviewer. Evaluations score every output with an LLM judge grading against the rubric you wrote — sourcing, thresholds, length, whatever “good” means for this job — and the scores accumulate into a trend.
When a score dips, you do not guess. Replay the run with overrides — a changed prompt, a different input — against the original as a controlled comparison, and ship the fix to the next tick.
evaluation — morning-report rubric
Stop babysitting “a script on a box somewhere.” Start operating “a schedule that fires exactly once, grades its own output, and files the evidence.”
What that unlocks
Mornings start finished
The 6am swarm ran, the rubric passed, the webhook fired, the artifact filed — all before the first person logs on.
Quality becomes a metric
Evaluation scores per run turn “is the automation still good?” into a chart — and a dip into a signal instead of a customer complaint.
Failures announce themselves
Signed, retried webhooks mean your systems hear about every completion and every failure. Nobody polls a dashboard to find out.
Debug by replay, not archaeology
Last Tuesday’s odd output re-runs with overrides against the original inputs. The diff answers in minutes what log-spelunking answers in days.
Keep exploring
Give your agent a workforce.
Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.