Skip to main content
Three scenarios showing how the primitives compose into real jobs, using what the platform does today (coming-soon pieces marked).

Nightly repo triage

A scheduled agent that wakes at 9:00 UTC, works, publishes a report, and sleeps. A deployment fires a fresh session on a cron tick; its computer clones the repo, runs the tests, and summarizes overnight changes into org-scoped Files.
Between fires, nothing is running — idle bills as storage. Each fire gets a fresh per-run budget; a run that would exceed the cap goes idle with stop_reason: "budget_paused", still reporting what it completed. The on_idle webhook pushes the result to you, and the published report stays in Files long after the fire.

A long report job on the cheap lane

A data-crunching run that takes hours and nobody is watching. An agent that delivers at nine has no reason to pay the immediate tariff — create the session with --window loose: same model, lowest price.
The event stream is keyed by sequence number, so it survives you: close your laptop, come back hours later, resume with --after-seq — no gaps, no duplicates. The durable runtime never depended on your connection anyway: wake, turn, commit, sleep. Send follow-up turns with vetta session send; the finished report lands in Files as a durable artifact.

An always-reachable ops agent

An agent that costs storage while it waits, and acts when the world calls. An ops agent holds an identityops@acme.com and a phone number. A customer texts the number; the SMS is carrier-verified and stored on the identity’s message feed. Your pager bot (or a person) starts a session from the stored message; the agent reads it, checks the runbook skill, acts through its connections under policy, and replies as the identity — then sleeps at storage cost until the next call.
The automatic wake — the stored inbound message starting the turn itself — is coming soon; outbound webhooks (session.idle, message.completed, budget.exceeded) are live for pushing results back to your systems.

Next: quickstart

Create an agent, run it, and read the meter in under five minutes.