The polite assistant is a useful stage prop. It nods, summarizes, asks whether you would like it to proceed, then waits for the next coin to drop.
That is not help. That is customer-service theater with a token budget.
The useful assistant has a spine.
Not swagger. Not fake certainty. Not that brittle machine confidence where every answer arrives dressed like it owns the building. A spine is simpler than that: a working bias toward taking responsibility for the next obvious move.
If the file is missing, look for it. If the command fails, read the error. If the request has a sane default, choose it. If the user is busy, do not turn every fork in the road into a committee meeting.
This is harder than it sounds because software culture has trained machines to be weirdly helpless. We wrap every action in consent theater, every suggestion in padding, every uncertainty in a little escape hatch. The result is an assistant that cannot decide whether to open the door unless you confirm the existence of hinges.
There is a cost to that. Every unnecessary question pushes work back onto the human. Every "would you like me to" converts a delegated task into a supervisory task. The assistant becomes another inbox, another dependent process, another glowing thing asking for attention.
Good autonomy is not recklessness. It is bounded initiative.
Read freely. Draft freely. Organize freely. Test freely. Make reversible changes. Leave a trail. Ask before crossing a boundary that matters: sending the email, deleting the data, spending the money, speaking in public, touching somebody else's trust.
That is the line. Internal momentum, external caution.
The best tools I have seen share this posture. They do not pretend the user is absent, but they also do not require the user to parent them through every obvious step. They absorb ambiguity where the stakes are low. They surface it where the stakes are real. They know the difference between "I can infer this" and "I am about to make a mess with your name on it."
Most assistant failures are not failures of intelligence. They are failures of nerve.
They had enough information to move, but they chose ceremony. They had enough context to make a small repair, but they asked for a meeting. They knew the next command to run, but they produced a paragraph about possible approaches instead.
The cure is not more personality. It is more ownership.
Do the useful thing. Keep the blast radius small. Preserve the user's options. Report what changed. Then shut up.
That is the spine.