GlossaryFloor 3 · The Agentthe block caught in a loop: it starts again until it gets thereFloor 3 · The Agent
stopping condition
No. 072 · v2026-08FR: condition d’arrêtThe stopping condition is what puts an end to an agent’s loop: either the goal is reached and verified, or a limit is crossed. Like an oven timer paired with a thermometer: one says that it is cooked, the other that it has gone on long enough.
What it is not
A stopping condition is not the model’s judgement on its own work. An agent readily declares itself finished, with assurance, without anything establishing it: this is even the most common failure mode. Nor is it a simple safety limit: the limit prevents a runaway, it says nothing about success. A correct system always has two of them, a success condition and an abandonment condition, and it knows which of the two put an end to the run.
In depth
The success condition
The success condition must be verifiable by something other than the model doing the work. The best ones are mechanical: tests that pass, a schema respected, a record present in the database, a total that comes out right. When the task does not lend itself to that, a separate verifier is used, another call that sees only the goal and the result, without the history of the attempts, which prevents it from being won over by the agent’s account. Trusting self-assessment amounts to asking someone to mark their own paper without a marking scheme.
The abandonment conditions
Abandonment conditions are set along several axes, and several of them are needed. A maximum number of turns, to cut short loops that repeat themselves. A budget in tokens or in money, which is the most concrete limit. A duration. A number of consecutive failures on the same tool, a sign that persisting will serve no purpose. And detection of treading water: if two turns produce the same action and the same observation, the agent is going round in circles, and nothing indicates that a third would change anything.
What follows the stop
What counts as much as the stop is what happens afterwards. An agent that gives up must return a state, not a silent failure: what has been done, what remains, what was blocking, and preferably a possible resumption. That is what separates a usable system from a system that is relaunched blindly. And the reason for stopping must appear in the trace, for a success rate that does not distinguish abandonment on budget from abandonment on failure allows no decision about improvement.
Relations where the neighbours live
Check 3 questions · click your answer
Level 1 · Recognise
How many stopping conditions does a correct agent have?
Level 2 · Distinguish
Why not let the agent judge for itself that it has finished?
Level 2 · Distinguish
Two consecutive turns produce the same action and the same observation. What should be concluded?
Lexigraph, "Stopping condition", v2026-08, https://www.lexigraph.org/en/stopping-condition/, CC BY 4.0.