What is an agent loop?

A task an AI agent re-runs on its own at a regular interval: every 5 minutes, every hour, every day. Where a classic agent works once toward a goal then stops, a loop starts over — to monitor, check or alert continuously, without human intervention.

How does /loop work in Claude Code?

You give a prompt and an interval (seconds, minutes, hours or days); Claude Code converts it to a schedule and re-runs the prompt at each tick, inside your working session. Built-in guardrail: a loop automatically expires after 3 days — it runs one last time, then deletes itself. A forgotten loop can therefore never run indefinitely.

What is the difference between a loop and a scheduled task (cron)?

A loop is ephemeral and session-scoped: perfect for “while I'm working” needs (watching a deployment, waiting for a result), with zero setup. A scheduled task — cron or GitHub Actions — is durable and runs without you, but requires real setup. Simple rule: need under 3 days → loop; permanent need → scheduled task.

What is a loop good for in retail?

Anything you check at regular intervals over a day: making sure morning orders move to fulfillment, tracking a critical restock, verifying a production deployment went well, checking that a supplier stock import landed. The loop looks, compares, and only pings you when there is something to say.

What is a “loop”? Recurring agent tasks

1. The loop: the floor above the agent

In What is an AI agent, I explained the assistant/agent difference: an assistant answers, an agent acts in a loop until the result is delivered — then stops. The loop adds the time dimension: the whole task is re-run at a regular interval. The agent no longer works once toward a goal; it comes back to check — again and again.

It's the natural pattern for anything monitoring: you never "finish" watching a deployment or an order flow; you look, compare against the expected state, alert if it drifts, and come back later.

Check → Compare → Alert if needed → Wait → Expire every interval, until expiry

The loop: a time-bounded monitoring cycle — not an infinite loop.

2. How it works in Claude Code

Claude Code (the tool I use for all dev work — see Chat, Cowork or Code) ships a /loop command: you give it a prompt and an interval (in seconds, minutes, hours or days), and it re-runs that prompt at every tick, inside your session.

The detail that won me over: the guardrail is built in. A loop automatically expires 3 days after creation — it runs one last time, then deletes itself. A forgotten loop can't run forever or silently burn a budget. It's exactly the kind of framing I called for in the agent guide (goal, budget, frequency, success criterion) — except here, the boundary is set by the tool itself.

3. Loop or scheduled task?

Both re-run a task at a regular interval; everything else differs.

Loop

Ephemeral, zero setup

Tied to your working session
Created in one line
Expires on its own (3 days max)

Scheduled task

Durable, real infra

Runs without you (cron, Actions)
Requires setup
Lives until someone stops it

Need under 3 days → loop. Permanent need → scheduled task.

My daily AI watch, for instance, is a scheduled task (it has to run every morning, forever). Watching that an Odoo deployment goes well this afternoon is a loop: useful for three hours, forgettable afterwards.

4. What it's good for in retail

Anything you check at regular intervals over a day or two:

Orders: check every 30 minutes that morning orders move to fulfillment; alert if a status gets stuck.
Production deploys: after a release, check every 10 minutes that key pages respond and nothing breaks.
Supplier flows: make sure an expected stock import arrived and integrated cleanly.
Critical restock: track an out-of-stock reference and ping as soon as stock comes back.

That's the workstream I'm opening at Cobra right now: moving checks I used to do by hand (or not at all) into loops. The detailed journal entry will land once the first ones run in production — as always, everything gets documented.