What a computer-using agent actually does all day

We gave a computer-using agent two weeks of real work — expense reconciliation, a device-refresh spreadsheet, and the weekly report nobody enjoys writing — and recorded every session.

The interesting part was not the failures. It was how often the agent got 80% of the way there and then needed a decision only a person could make: which of two inconsistent numbers to trust, whether a refund was worth chasing, when to stop.

Where the time actually went

Three patterns showed up often enough to name:

• Re-reading its own output. Given a long spreadsheet, the agent would re-derive context it had already established, burning tokens on orientation.
• Silent assumptions. It picked a convention and stuck to it without flagging the choice, which is fine until the convention is wrong.
• No sense of "good enough". It would keep polishing a report that was already finished.

What it cost

Cheaper than an afternoon of my time, more expensive than I expected, and the variance between tasks was wider than the average is useful for.

We close with the week in news: two model releases worth caring about, one procurement story that will matter more than either, and a governance draft that is quietly sensible.