Answer
How do you stop an AI agent from stopping too early?
Give it a completion condition it can test and a coverage count it must report. Encouragement does not change stopping behaviour.
Give it a condition it can test rather than judge, and require it to report how many items it covered against how many exist. Premature stopping is a measurement problem, and instructions to be thorough do not address it.
Early stopping has a specific cause, and it is not laziness or a shortage of effort. The agent halts when its assessment says the goal has been met, and where no testable goal was supplied that assessment is a judgement about resemblance. Enough work has been done that the situation looks like a completed one. Since producing things that look completed is what the system does best, the assessment is systematically optimistic, and it will be optimistic no matter how firmly it was asked to be thorough.
The first control is the completion condition, stated before the run and expressed as something checkable. Not do a thorough job but every record has a category and no record has the category unknown. The second phrasing has a failure state the agent can detect. This alone removes a large share of early stops, because the loop now has something to compare against other than its impression of progress.
The second control is coverage accounting, and it addresses the case the first one misses. Require the agent to state how many items exist, how many it handled, how many it skipped and why. This is effective for a slightly unobvious reason: the count forces the agent to look at the whole set, and looking at the whole set is exactly the operation it otherwise skips. An agent that has never enumerated the input cannot know it stopped early, so it reports completion truthfully as far as it knows.
The third is preventing the definitional retreat, which is the most irritating version of this failure. Asked to fix all the failing cases, an agent will sometimes fix several and then explain that the remainder were out of scope, misconfigured or pre-existing. Some of those explanations are correct, which is what makes the pattern hard to police. The control is to require the disposition of every item individually, so a claim of out of scope is attached to a named item and can be checked, rather than applied as a blanket to whatever is left.
The fourth is structural: long tasks with no intermediate landmarks stop earlier than the same work divided into stages with an output each. This is not a motivation effect. A stage boundary is a natural place for an assessment of completion, and an assessment against a small stated stage is far more accurate than one against a large vague goal. Dividing the work is therefore a completion control as well as a context control.
A caution worth stating. Every one of these makes stopping harder, and the opposite failure is real: an agent that cannot stop will keep working, keep spending and keep changing things past the point of usefulness. The controls above should be paired with an explicit statement of what is not in the job, or you have traded one failure for a more expensive one.
An agent stops when the work resembles finished work, so the fix is to make finished a thing that can be counted rather than recognised.
Siddharth Sharma, Context Theory
Related questions
Does asking it to double-check help?
It produces a second pass over the work already done, which is not where the missing work is. Reviewing what was completed cannot surface what was never attempted, and the review will come back positive because the reviewed portion was fine. Coverage accounting works where double-checking does not, because it operates on the input set rather than on the output.
Is early stopping worse in long sessions?
It is more likely, and for an identifiable reason: as the session grows, the original statement of the job is a smaller fraction of what the system is attending to, and the recent work is a larger one. Restating the completion condition at intervals is cheap and measurably effective, and it is the same remedy as for drift generally.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Agents who give up after one contact | 44% | Category-wide |
2026 speed-to-lead benchmark · verified
Multi-study aggregate · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Premature stopping arises from an assessment of resemblance rather than from insufficient effort, so it is systematically optimistic in a system optimised to produce work that resembles completed work, and instructions to be thorough do not reach the mechanism. | Comparing stopping points across runs of the same task with and without an emphatic thoroughness instruction. |
| Software | Requiring a coverage count of items existing, handled and skipped forces enumeration of the input set, which is the operation an agent otherwise omits and without which it cannot detect that it stopped early. | Comparing the item count in an agent's coverage report against the count in the source set. |
| Response | Definitional retreat — declaring remaining items out of scope, misconfigured or pre-existing — is policed by requiring a disposition for each named item, because a blanket exclusion cannot be checked while a per-item claim can. | Requiring a per-item disposition list and checking a sample of the exclusions against the items themselves. |
| Buying behaviour | Dividing long work into stages with an output each improves completion accuracy because an assessment against a small stated stage is more reliable than one against a large vague goal, which makes staging a completion control rather than only a context control. | Running the same body of work as one task and as staged tasks, and comparing coverage in each. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one