Agent Loops

Most "AI workflows" on the internet are a prompt you have to remember to run. An agent loop is different.

It runs whether you remember or not.

This is, in Anthropic engineer Boris Cherny's phrasing, "the simplest thing that works." It's also most of what a solopreneur actually needs.

How it works

``


Schedule: every Monday 8am (or hourly, or on-event)

Trigger: cron / scheduler kicks the agent

Agent: reads inputs, runs the prompt, produces output

Delivery: output lands in Slack / email / Notion / a draft

Human: reviews and approves — does not initiate

``

The shift is small but it's everything: you become the reviewer, not the originator.

Most of your work week is initiation. "What was I going to do next?" "Did I send that?" "When was the last time I…?" Loops eliminate that whole layer. The work shows up. You either approve it or you don't.

What makes a good loop

• Same shape every time. If the inputs vary wildly, it's not a loop yet — it's a request for help.
• Tolerable when imperfect. A loop that drafts a recap you edit is fine. A loop that auto-sends customer messages without review is not — yet. (The trust ladder decides which version you ship.)
• Saves at least 30 min/week. Below that, the cost of running and reviewing it eats the savings.

What loops replace

• Weekly client recap drafts
• Monday morning planning prep
• Inbox triage and tagging
• Lead-list enrichment
• Daily journal / metrics summary
• Recontact reminders for old leads

Each is a recurring solopreneur tax. Each is a one-to-two-hour build and a two-to-five-hour return — every week, forever.

Why this site emphasizes loops

Most bottlenecks on this site are best fixed as loops, not one-shot prompts.

A prompt is a tool. A loop is an employee that doesn't need onboarding.

If you can't see how a fix becomes a loop, you haven't found the real bottleneck yet. You've found a chore that wants attention. Maybe the discipline is to not build it as a workflow until you can see the schedule.