Try it

Run a loop. Watch it stop for an honest reason.

Most of loop engineering is invisible until you feel it: the budget draining on every retry, a verifier that’s too loose waving a bug through, the moment a run gives up and asks for a human. Set the dials, press Run, and watch the token travel the seven‐step lifecycle — trigger, discover, delegate, act, verify, persist, decide — turn after turn. Nothing here calls a model; it is the same arithmetic the runnable example uses, drawn so you can see it turn.

Loop idle. Choose a scenario or press Run.

Scenarios

Stop philosophy

/goal halts the instant a separate checker confirms the done‐condition. The workhorse of serious loops.

Task difficulty0.35 Harder work improves more slowly each retry. Verifier strictness0.80 Below the dashed line on the quality bar, the checker passes work that is not actually correct. Max retries6 The iteration ceiling. Hit it without a pass and the loop stops. Token budget30,000 Each delegate costs more than the last: the run re‐bills its growing history. That is the O(N²) tax. Flaky verifier — the checker’s verdict gets noisy, so it sometimes fails good work or passes bad.

speed

loop‐receipts.log

What you just watched

Every run ends one of five ways

Four are honest stops; the fifth is the one that should scare you. A loop you can walk away from is one whose worst ending you have already seen.

honest stop
Goal met. The separate checker confirmed the done‐condition and the work was genuinely correct. The only ending you actually want.
honest stop
Budget spent. Tokens ran out before the goal. The machine clock fails loudly, so you can bound it with a number. Token economics →
honest stop
Stalled. The same failure kept repeating with no new evidence. Stopping beats thrashing.
honest stop
Needs a human. The change was borderline and the checker would not certify it, so it escalated. This is a success state, not a failure.
the dangerous one
Shipped a bug. The checker passed work that was not correct, because its strictness sat below the true bar. A verifier weaker than the work is worse than none: it manufactures false confidence. Why the checker is the asset →

Notice what /loop does that /goal never would: it keeps spending after the work is already correct, because a cadence does not know whether it succeeded. That gap is the whole reason /goal exists. Read the chapter →

Learn it in depth Build a loop for your workflow