AI loop
A bounded operating cycle where an AI system repeatedly acts, checks evidence, and decides whether to continue, stop, or ask for approval.
Glossary
AI loop glossary
Short, operator-grade definitions for the terms around AI loops and agentic workflows.
DetailsTerms
Loop engineering
Designing the trigger, objective, verifier, stop condition, and approval boundary around an AI agent.
Loop stack
A layered model of repeated AI work, from token loops and tool loops to verification, goal, scheduled, and operating loops.
Stop condition
The explicit rule that ends a loop: success, blocker, budget, risk, or required human approval.
Verifier
The evidence check that decides whether a loop’s output is acceptable.
Approval gate
A boundary where a human must approve before the loop publishes, sends, deletes, spends, shares, or changes accounts.
Heartbeat loop
A recurring check that runs at a steady interval to notice changes or maintain state.
Cron loop
A scheduled loop that runs on a calendar or interval.
Goal loop
A loop that keeps working toward a stated outcome until the verifier passes or a stop condition fires.
Deterministic loop
A loop whose success can be checked by an external fact such as a test, metric, file diff, source check, or HTTP response.
LLM-as-judge loop
A loop where the model grades quality against a rubric; useful for subjective work, but more brittle than deterministic verification.
Completion promise
A done contract that prevents an agent from stopping until every acceptance criterion is proven or explicitly blocked.
Loop budget
The maximum time, tokens, dollars, files, retries, or risk a loop can spend before it must stop or ask for approval.
Bounded repeat
A repeated action constrained by scope, budget, step limit, source of truth, or approval boundary.