AI agents
Persist model and tool boundaries, enforce budgets, and wait safely for approval.
Use cases
The same engine can serve different industries, but the useful design question stays concrete: what must survive a crash, how long can it wait, and which effects must never be repeated?
Persist model and tool boundaries, enforce budgets, and wait safely for approval.
Run onboarding, billing, lifecycle messaging, and account tasks across retries.
Make consequential effects explicit, idempotent, and reconcilable.
Coordinate scheduled steps and replies without duplicate sends.
Separate chain confirmation from durable off-chain process state.
| Process trait | Likely starting point |
|---|---|
| One short scheduled task | Cron with overlap protection and alerts |
| Many independent jobs | A queue with idempotent consumers |
| Multi-step state across failures or days | A durable workflow |
| A consequential external call | An effect ledger and reconciliation, whatever the scheduler |
The documentation is the source of truth for current behavior.