An assistant that answers well is useful. An assistant that opens the file, makes the change, runs the test, sees it fail and tries again is a different category of tool, and that is where development is concentrated.
What makes it work
A feedback signal. Agents perform well precisely where success can be measured automatically – a test passes or does not, a value is extracted or is not, a form validates or does not.
Give the same system a task with no measurable outcome and it produces something plausible with no way to know whether it is right, which is worse than useless because it looks finished.
The permissions question
An agent that can act needs access, and access is where the risk lives. The pattern that will settle is tiered: read freely, write with logging, and destroy or spend only with explicit human confirmation.
Organisations that skip that structure will learn it the expensive way, and some already have.
What gets automated first
Work that is repetitive, has a clear definition of done, and where errors are cheap to detect. Data entry, migration, format conversion, routine code changes, first-pass document review.
Not judgement, not negotiation, not anything where the right answer depends on context nobody wrote down.
The realistic timeline
Narrow agents in production now, doing specific jobs under supervision. General-purpose agents that handle open-ended work reliably are further off than the demonstrations suggest, because the demonstrations are chosen to have clear success criteria.
What to do about it
Identify the one task in your work that is repetitive and verifiable. Automate that. Measure the time saved including the checking. That single exercise teaches more than any amount of forecasting.