Loops all the way up

Verify one run ends with one small, real thing: a loop with one verified pass and a log to show for it. This guide is about what that loop can become.

A verified loop can become one pass inside a bigger loop. The bigger loop treats the whole inner loop the way the inner loop treats a single action: run it, check the result, and decide whether to continue, stop, or escalate. Loops stack this way, and the layers have names.

Three layers, named by what varies

Each layer is named for the one thing it varies between passes.

Layer	Varies between passes	The check looks at	Stops when
Task loop	Actions	One task's declared check: tests green, link resolves, file exists	done when the check passes; blocked, exhausted, or interrupted when it cannot complete within the written rule
Goal loop	Tasks: which task loops to open, retry, or drop	The goal's stopping condition	done when the goal is met; blocked, exhausted, or interrupted when the task sequence cannot continue
Steering loop	Goals	World state: metrics, external changes, qualitative aims	Rarely reaches done; usually stops as blocked, exhausted, or interrupted at human decision points

The loop you verified in the last guide is a task loop: it varies its actions until one task's check passes. A coding agent that edits, runs the tests, reads the failures, and edits again is the everyday example.

A goal loop sits above it and varies the tasks. Working a milestone is the everyday shape: pick a task, run its task loop to green, pick the next, until the goal's stopping condition says done, blocked, exhausted, or interrupted.

A steering loop sits above that and varies the goals themselves. It watches real-world state: metrics, external changes, and qualitative aims. Its job is to decide what the goal loops should pursue next, and to pause at the right moments so a human makes the calls that matter.

The same five parts at every layer

Nothing new is invented on the way up. Every layer must answer the same five questions from Loop or not?: a repeated pass, carried state, an explicit check, a written stop or escalation rule, and a harness that owns the repetition and records what happened. Only two things change with the layer: the scope of what varies between passes, and where the check looks.

Two properties follow, and both get violated in the wild:

• A higher layer is not more loop than a lower one. A steering loop with no check is not a loop at all, while a task loop with a verified pass is the real thing.
• Nothing is inherited. Checks, stop rules, and trust do not flow up or down the stack. Each layer needs its own five parts, written down at that layer.

Only stack verified loops

Trust climbs the stack from the bottom, one verified pass at a time.

• Base case. A task loop is trusted when one verified pass exists. That is the whole job of Verify one run.
• Induction step. A loop may vary only loops that are themselves verified. A goal loop may orchestrate only task loops that each have a verified pass. A steering loop may re-aim only verified goal loops. And the stacking loop needs its own verified pass, with its own check, stop rule, and log, before anything stacks on it in turn.

The consequence worth memorizing: a stack is never verified top-down. There is no such thing as a verified steering loop running over unverified task loops. If the bottom has no pass log, the top is a plan, not a loop.

Where the ladder stops and the layers start

Loop or not? climbs a ladder: prompt, scheduled prompt, automation, workflow, agent workflow, loop, agent loop, harness, autonomous system. The ladder and the layers answer different questions, and they do not compete.

• The ladder classifies decision and control structure. It answers: is this a loop at all, or a prompt, a schedule, a workflow?
• The layers classify nesting scope. Once something qualifies as a loop, they answer: what varies between its passes, actions, tasks, or goals?

So layers apply only at and above the ladder's loop rung. A scheduled prompt is not a task loop, because it is not a loop. Run the five-question test first; reach for a layer name second.

What is proven here, and what is not

Layer names are cheap to say and expensive to earn, so hold this page to the site's own standard.

As of this writing, the pass logs published on this site include task-loop receipts and one bounded goal-loop receipt: loops/log/2026-07-10-first-goal-loop-receipt.md. That receipt proves only the specific goal-loop pass it records. There is still no checked steering-loop pass log here, so this guide defines steering loops without claiming loopmaster.ai operates one. Steering loops in particular are mostly unproven everywhere, so treat any claim of one, from anyone, as unverified until you see the log.

The bar does not move as you go up. Claim a layer only with that layer's own pass log: what ran, what the check found, what state was written, and how it stopped. Loops go all the way up only as far as the evidence does.