what a loop actually is
a goal with a finish line, context, an action, a look at the result, an adjustment, repeat. one-shot generation cannot see its own failure. the loop can, and that is the whole point.
but a loop is only a loop if the check is real. if the only verifier is the model grading its own output, you don't have a loop. you have a model complimenting itself with extra steps. the check has to live outside the model: a test suite, a typecheck, a linter, a human at a gate. strip that out and it stops being a loop and becomes a runaway.
what separates a loop from a runaway
five parts. miss the last two and you have a token furnace nobody is watching:
- a done-when. a hard finish line the agent can test against. no target, no loop, just motion.
- external verification. "done" is a claim that needs evidence from outside the model making it.
- stopping rules. max iterations, a cost cap, or done. every pass is a full model call plus context, so loops multiply spend quietly. uncapped, that is the difference between an agent and a bill.
- memory. it records what it already tried, so it does not circle the same wall every pass.
- a heartbeat. it emits a pulse so you can see it is alive. an agent that can go quiet without anyone noticing is not finished. it is missing.
the monitoring is part of the build
judge an agent's health against its own cadence, never a flat threshold. a five-minute poller silent for six hours is broken. a weekly job silent for six days is fine. one flat "last seen" rule gets this wrong in both directions: false alarms on the slow jobs, real silence on the fast ones.
and an alert has to carry the action. severity is not information; the next physical move is. "urgent: engine degraded" with no move attached is inbox noise, and noise trains the operator to ignore the one alert that mattered. so the rule in our builds: green stays silent, red arrives with the move already drafted, or it says "no action needed" and means it. you wake to signal, not noise.
last, the boring one that saves you: an unsupervised loop never runs on a live main branch. branch it, cap it, gate it. an agent guessing past a gate on a client repo is how a small bug becomes a story.
why this is the actual deliverable
the model is rented. every studio runs the same weights you do. the loop is the part you own, and on identical weights it swings the result more than any model upgrade will. when a client commissions an ai agent, the agent is the easy half. the loop around it, the verification, the stopping rules, the heartbeat, the alert that tells you what to do, is the product.