Metaorcha

Agent loop

Agent loops repeat the work until the check passes.

A loop runs the same task again and again. Each pass ends with a check you name, such as a test command. When the check passes, the loop stops. When it fails, the loop either climbs to a stronger engine or stops, depending on the setting you chose.

Each pass writes its own receipt. A loop with twelve passes leaves twelve receipts, and each one checks on its own.

MetaBot has three loop shapes today

  • Ladder: ask --ladder starts with the cheapest engine. It climbs to the next engine when a failure is recorded.
  • Check: --check CMD judges each rung of work. A failing check climbs the ladder.
  • Routine: a goal that runs on a schedule, with no person there to answer. Routines use the same approval rules as a run.

What you can check yourself

  • A ladder run on a planted bug: the cheap engine diagnosed it and made no edit, which the receipt shows.
  • A scheduled run with no person present: a destructive call stopped, and the receipt records the stop.

Claim pending: the proof rows behind these checks are not linked from this page yet.

Questions

What is the difference between a loop and a graph?
A loop repeats one task until a check passes. A graph runs different steps in the order their dependencies allow.
How is a loop different from running the task again by hand?
The loop applies the same approval rules on each pass and leaves a receipt for each pass, so you can read what each attempt did.