GlossaryFloor 3 · The Agentthe block caught in a loop: it starts again until it gets thereFloor 3 · The Agent
goal
No. 068 · v2026-08FR: objectifA goal is the outcome an agent has to reach, stated once and held for the whole loop, whatever steps it takes to get there. Like an order placed with a craftsman: you say what you want to end up with, not the sequence of movements.
What it is not
A goal is not an instruction. An instruction says what to do and is consumed in one go; a goal says how you will recognise that the work is finished, and it survives every intermediate attempt. Nor is it an intention held by the system: the agent wants nothing, the goal is a piece of text that the harness re-injects at every turn so the model always knows where it is heading. Without that re-injection, nothing of it would remain.
In depth
What makes the agent
It is the goal that makes the difference between an agent and a workflow. In a workflow, a person wrote the sequence of steps; in an agent, they wrote the end to reach, and the model decides the next step at every turn. All the flexibility and all the unpredictability come from there. This carries a consequence that is rarely anticipated: the quality of an agent depends first on the quality of the statement, well before the model used.
The three elements
A good goal carries three things, and it is usually the third that is missing. The expected outcome, described by what will exist at the end: a ticket booked, a note published, a case file completed. The bounds, meaning what is forbidden, what requires human approval, what budget and how many attempts. And the stopping condition: how one recognises that the work is done, phrased so that the system itself can check it. Without that last point, an agent runs until its budget is exhausted, or declares victory too early.
Two drifts
Two failures lie in wait, and they are symmetrical. A goal that is too broad, of the “improve our customer relations” kind, has no observable end: the agent wanders an infinite space and produces decorative work. A goal that is too narrow, one that in fact describes a sequence of steps, did not call for an agent: a workflow would have been more predictable, cheaper and easier to guarantee. Choosing an agent means agreeing to pay in unpredictability for adaptability, and that is only justified when the possible paths really do vary.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
What does not change while an agent is working?
Level 2 · Distinguish
Which element is most often missing from a badly stated goal?
Level 2 · Distinguish
You hand an agent a task whose steps are always the same. Good choice?
Lexigraph, "Goal", v2026-08, https://www.lexigraph.org/en/goal/, CC BY 4.0.