How many things should an agent do at once?

The unit is not the number of steps or tools but the number of claims the output makes. A run that reconciled the records and also tidied the naming has produced two results with one review attached, and the reviewer must accept the tidying to get the reconciliation or reject the reconciliation to refuse the tidying. This is the practical cost of unfocused runs and it appears at review rather than during execution.

The same argument applies to the multi-part task that arrives as one request. Handle the enquiries and update the report and check the outstanding items is three jobs, and running them together produces a report mixing three sets of findings, three sets of failures and a coverage figure that means nothing. Splitting them costs three briefs and produces three things that can each be judged.

There is a genuine efficiency argument for combining, and it is smaller than it appears. Work that shares context — the same records, the same investigation, the same setup — is cheaper done together, and that saving is real. What it buys is tokens and time; what it costs is reviewability, and reviewability is usually the constraint. Combining is right where the outputs will be accepted or rejected together anyway, which is a narrower set than it sounds.

Within a single run, doing several things simultaneously is a different question with a simpler answer. Parallel tool calls that gather independent information are straightforwardly good — reading four files at once rather than in sequence costs nothing extra and saves time. What should not be parallel is anything where one result should inform the next, which is a dependency mistaken for an opportunity.

The signal that a run did too much is available before any reading: the size of the result. A change touching eleven files for a task that should have touched two, a report covering four subjects, a summary with three unrelated sections. Each of these is visible immediately and is more reliable than any assessment of whether the work was appropriate, because it does not require knowing what the work was.

One qualifier. A run that discovers something outside its task should report it, not ignore it, and reporting is not doing. The discipline is that the additional finding arrives as information while the run's output remains a single claim, which preserves both the reviewability and the value of having noticed.