Lexicon · III · Orchestration
Agentic loop
Also: loop
The cycle that turns AI from talk into action: act, observe the result, correct, repeat, until the goal is reached or the blockage is acknowledged.
The analogy
Parallel parking in a tight spot. Nobody computes the full trajectory before touching the wheel: you turn, look, correct, repeat, and the car ends up parked. It is a loop of action and observation, not a perfect plan executed in one stroke. The agentic loop applies exactly that principle to work entrusted to AI.
One mission, turn by turn
Step through the turns: nobody dictates the steps; what was just observed decides what comes next.
The mission
“Find out why invoice 4021 was rejected.”
Harness guardrail: 8 turns maximum
This is all the agent receives: a goal, not a procedure. Start the loop.
Five turns, five inferences, one real cause, with not a single step dictated.
The anatomy of a turn
Each turn chains three moments. Decide: one inference; the model rereads everything in front of it, including the observations of previous turns, and chooses the next action. Act: a tool is used to read a file, query a system, write. Observe: the result of the action is fed back into the context window, where it will inform the next decision. That third moment is what makes everything work: the loop turns error into information. And every turn leaves a trace. The turn log is the first object oversight must know how to read.
A concrete example
An agent must fix a bug: it edits the code (action), runs the tests (observation), sees three fail (reality pushing back), adjusts its fix, reruns. Four turns of the loop later, the tests pass. Each turn combines one inference with one tool-assisted action, and the log of those turns is precisely what human oversight must be able to read.
When the loop goes wrong
Three typical failures. Stubbornness: the loop repeats a fix that does not work, barely rephrasing it; each turn confirms the previous one instead of questioning it. Drift: it reaches a sub-goal, then invents a follow-up nobody asked for. Exhaustion: it spins, every turn costs an inference, and the bill runs with no result. None of these failures is fixed "in the model": they are prevented in the harness, with turn limits, per-mission budgets, mandatory stopping points and a human able to read the log and pull the plug. A loop without limits is not autonomy: it is delegation without a mandate.
What it is NOT
It is not mere repetition: every turn re-decides what comes next based on reality; that is where the "intelligence" lives. And it is not infallible: a loop can dig into a dead end or spin forever. Hence the harness's guardrails: turn limits, budgets, mandatory stopping points.
Often confused with
The distinction that changes the conversation.
- Classic automation (RPA, scripts)
- A script replays a fixed sequence and breaks at the first surprise. The agentic loop decides the next step based on what it has just observed. One executes a pre-drawn path, the other pursues a goal, which makes it more powerful, and less predictable.