What is Agent Loop?

The loop is what separates an agent from a single model call. One turn in, one answer out becomes: look at the goal and what's known, choose the next action, execute it, fold the result into state, repeat. A booking agent might loop six times — check the calendar, find a slot, look up an address, draft an invite, hit approval, send. Loops need exits: a completion condition, a step or token budget, and a rule for what to do when a tool keeps failing.

Velaris runs agents as LangGraph graphs, which makes the loop a real structure rather than a while-statement around a chat call: state is explicit, steps are nodes, and a run can suspend and resume. That's what allows an approval to interrupt mid-loop — a destructive-tier tool call halts the graph, waits for a human decision, then continues from exactly where it stopped. Credentials are scoped to the run, so the loop's reach is bounded no matter how many iterations it takes.