An agent's time limit cannot tell you whether its work is finished
Define an inspectable result before extending an agent's runtime, and read time-horizon benchmarks as task measurements.

An agent can spend four hours pursuing the wrong interpretation of a task. It can also finish a difficult task in minutes. A time limit helps contain the first case, but it tells you little about whether the result is correct.
The useful limit is a piece of work whose outcome you can check. I would give the agent one bounded deliverable, define the evidence needed to accept it, then decide which actions that deliverable requires.
A benchmark hour is a measure of task difficulty
METR's time-horizon measurements use the time a human expert needs to complete a task. A model's 50% horizon describes predicted success at that level of task difficulty. It does not mean the model can safely act alone for that many hours.
The distinction is easy to lose when a chart's axis is labelled in minutes or hours. A two-hour task in that evaluation may take the agent a different amount of wall-clock time. The benchmark also covers a particular task suite and agent setup.
As of November 15, 2026, the public chart says it was last updated on September 8 and is no longer actively maintained. Treat it as a dated measurement. It cannot establish a universal 45-minute rule for today's models.
METR's analysis of modelling assumptions also shows that fitted horizons depend on methodological choices. That is another reason to read the measurement definition before turning a number into an operating policy.
Give the run an observable finish
Consider a hypothetical repository maintenance task. "Improve test coverage" leaves the agent room to keep expanding its work. "Cover these two parsing branches, keep the public interface unchanged and present a reviewable diff" supplies a finish that someone can inspect.
The second task still needs judgement. A test can execute a branch without detecting a meaningful defect. Acceptance should include examples of the behavior that must be preserved, along with a failure case that the new test should catch.
Choose the deliverable before setting the time budget. If a useful result requires changing unrelated modules, the agent should surface that dependency. It should not reinterpret a small maintenance request as permission for a redesign.
Intermediate checkpoints can preserve useful work. A completed investigation may justify a proposed patch even when the implementation needs another session. Record the evidence and the unresolved question so the next session can resume without guessing.
Separate completion from the next permission
A generated patch, a successful check and a deployment are different events. Define which of them the user has authorized. The agent should be able to finish the permitted work without treating every successful step as permission for the following one.
For an internal data cleanup, the accepted output might be a proposed change set with uncertain matches marked for review. For a production correction, existing authorization may include applying the change and verifying the live behavior. The task's scope determines that boundary.
This is more useful than declaring every public action forbidden. Some public actions are the requested outcome. The control is whether the action has a defined target and a way to establish that it succeeded within the authorized scope.
Make ambiguous external results a stop condition. If an API times out after a write, establish whether it committed before attempting another write. Continuing the loop with the same uncertainty can multiply the damage.
Use budgets to detect a stuck run
A runtime cap limits elapsed time. A tool-call budget limits repeated attempts. A spending cap limits a different resource again. Set the constraints relevant to the task, then report which one ended the run.
A status update should contain evidence of progress. "Still working" is weaker than "the reproduction passes; the new assertion fails on the original code." That observation helps a reviewer distinguish useful investigation from repeated attempts at the same failed approach.
Anthropic's long-running agent harness report describes progress records and feature checks used across context windows. Those are concrete design ideas from a specific implementation. They do not remove the need to define success for a different workload.
Try the proposed boundary on a small sample of your own tasks. Record accepted results and the interventions each needed. Keep failures in the denominator, including runs that ended without a deliverable.
That record gives you a basis for extending autonomy. If the agent repeatedly completes a defined task and detects its own relevant failure conditions, a larger assignment may be justified. If its status reports keep expanding while the acceptance evidence stays empty, more runtime is unlikely to resolve the underlying specification.