Nightschool AI: Build your first self-running agent loop
A self-running agent loop is a small setup that runs an AI coding or agent tool on a repeating schedule instead of only while you're watching it: it reads a plain-text state file to pick up where it left off, does one step of work, has a separate process check that step before anyone trusts it, and stops to ask a person when a decision needs a human, such as spending money, sending something, or deleting something.
What you'll have built
- A scheduled loop that wakes up on its own and runs one step of work without you at your computer
- A state file the loop reads and updates every run, so it always knows what it already did and what's next
- A separate verifier step that checks the work before it's trusted, because nothing should grade its own homework
- An escalation step that stops and asks a person, with a recommendation, whenever the task needs money spent, a message sent, or something deleted
What's inside
Guide
Setting up the schedule, writing the state file, and wiring in a separate verifier and an escalation step, in that order, with a real operator's own failures explained so you can plan around them first.
Copy-paste templates
A copy-paste state-file format, a verifier checklist template, and an escalation message template that states the gate it hit and its recommendation.
Test checklist
How to prove the loop actually resumed from its state file instead of starting over, and that the verifier catches a run you deliberately break.
Worked example
A walk-through of a real loop's rough edges, abstracted: a run that kept its machine busy until it froze, a usage limit that stalled work overnight, and a wake-up schedule that quietly stopped firing until a heartbeat check replaced it.
Not ready to join the waitlist? Start with a free lesson instead: Hooks and automation (free lesson)
Frequently Asked Questions
Do I need to know how to code?
No. The kit is written for the never-coded-before level: every command and config file is shown and explained, and the loop itself is built from plain-text files, not a programming language.
What exactly stops it from doing something risky on its own?
A short list of hard gates you set before it ever runs. Spending money, sending a message, or deleting something always stops the loop and hands the decision to a person, with a recommendation attached, rather than letting the loop decide for itself.
What goes wrong in practice?
In the real loop this kit is based on: a run that kept its machine busy until it froze, a usage limit that stalled work overnight with no one watching, and a wake-up schedule that silently stopped firing until it was replaced with a heartbeat check. The kit walks through each one and how to catch it sooner.
Is this tied to one AI coding tool?
No. The loop design works with any AI coding or agent tool you can run from a command line, such as Anthropic's Claude Code, OpenAI's Codex, or GitHub's Copilot. Nightschool AI is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, or GitHub; they're named only to describe tools the loop can drive.
When does the kit ship?
It doesn't exist yet. We're measuring demand first: join the waitlist and tell us what you'd want it to do.