The Agentic Loop: Stop Conditions, Verifiers, and Why Agents Don't Know When to Stop
Most agent failures are loop failures. Design stop conditions, verifiers, and retry budgets that make LLM agents finish reliably instead of wandering.
A loop is one decision made over and over: call again, or stop. Almost every long-running agent failure reduces to that decision being made by nothing in particular — the model is asked whether it is done, it says yes, and no one checks.
Why agents overrun
An agent without a verifier optimises for the appearance of completion, because that is the only signal in the loop. Give the loop an external check and the same model, on the same task suite, stops in a different place.
01 · Stop conditions
Written down, not implied
If done-ness is not expressible as a check some other process can run, the loop has no stop condition — it has a vibe.
02 · Verifiers
Cheap and external
A test run, a [schema parse](/posts/tool-design-for-agents/), a diff against expected state. The weakest external verifier outperforms the strongest self-report.
03 · Budgets
Retries are a resource
Bound the loop in steps and in tokens, and make exhaustion a distinct, visible outcome rather than a silent partial success.
"An agent that cannot fail loudly will fail quietly, on your users, at scale."
The experiment: one task suite, retry budget held fixed, verifier strength as the single variable. Traces published with the result.