How loops work
The shape every 1mn feature fits — trigger, discover, decide, act, verify, gate, persist, hand off — and the three properties that decide how autonomous it can be.
A loop is a system that, on a schedule or trigger, discovers work, hands it to an agent, verifies the result, records what happened, and decides the next move — until the goal is met or it hands back to you.
The shape
Every loop in 1mn fits one shape:
| Stage | What happens |
|---|---|
| Trigger | A cron, a webhook, an event, or you. |
| Discover | Pull the relevant signal (errors, metrics, drafts) and dedupe against what’s already handled. |
| Decide | Filter to what this loop is allowed to attempt right now; skip the rest. |
| Act | An agent run produces an artifact — a PR, a draft, a campaign, a report. |
| Verify | Prove it’s correct — tests, a critic, a metric check. This is the hard part. |
| Gate | Irreversible / high-impact actions pause for you. Reversible ones auto-apply. |
| Persist | Record what was tried and the outcome, so it isn’t re-done next time. |
| Hand off | Surface to you only at the edge — low confidence, repeated failure, or a gate. |
What decides how autonomous a loop can be
Three properties of the work (not the agent):
- Verifiability — can the loop prove it succeeded? A clean success signal (tests pass) buys autonomy; “looks right” doesn’t.
- Blast radius — is a mistake reversible? Wrong docs are recoverable; a bad prod deploy or a sent email isn’t.
- Ambiguity — one right answer or fifty? Patching a known bug has one; “build a feature” has a hundred.
This is why engineering loops mature first (tests give a clean signal) while money and strategy stay human-owned.
The trust ladder
Most loops ship in draft-only or recommend mode and graduate one specific action at a time, only after the verify stage has earned it. You decide how far each loop is allowed to go from Cadence — turn one on, watch what it produces, and loosen the gate when you trust it.