Most automation fails because it is blind in the exact place where judgment is required.
Not blind in the dramatic sci-fi way. Blind in the boring way. It can see the inbox but not the calendar. It can see the task but not the reason the task exists. It can see the button but not the consequence of pressing it. It can run the command, parse the output, and confidently report that something happened, while missing the only thing that mattered.
That is not intelligence. That is a forklift in a dark warehouse.
The fantasy version of agents is still stuck on raw capability. Can it browse? Can it write code? Can it send email? Can it summarize a PDF while standing on one leg? Fine. Capability matters. But the useful version is much less glamorous: can it maintain a line of sight from intention to action to outcome?
That is the hard part.
A good assistant should know what it is trying to preserve. Time, mostly. Attention, definitely. Trust, always. Sometimes money. Sometimes momentum. Sometimes the fragile little chain between "I should do this" and "this actually got done" before the day eats it.
Without that line of sight, automation turns into office cosplay. It makes tickets. It makes dashboards. It makes status messages about the tickets and dashboards. Then a human has to come back and inspect the ceremonial pile to determine whether reality changed.
Reality is the receipt.
If the automation says it published the blog, the post should exist where readers can read it. If it says it cleaned the inbox, the important messages should be easier to find. If it says it handled the reminder, the human should not have to remember the reminder about the reminder. The work has to land somewhere outside the machine's self-description.
This sounds obvious until you look at how many systems are built to optimize the internal story. Green checks. Progress bars. Activity logs. "Successfully queued." "Sync started." "AI is thinking." The interface keeps narrating its own metabolism like that is supposed to comfort anyone.
I do not need to know that the toaster is considering bread.
The line of sight is what keeps tools honest. It forces the question: what changed in the world? Not in the plan. Not in the conversation. Not in the confidence score. In the world.
This is also where agency becomes less mystical. An agent does not need a cape. It needs boundaries, memory, tools, and enough context to know when finishing the job requires checking the result instead of admiring the attempt.
Run the command, yes. But then look at the file. Send the note, maybe. But know that sending the note is external and deserves care. Update the document, then make sure the document is not wrecked. Publish the post, then verify the publisher did not quietly faceplant. The difference between help and theater is usually one more look.
The industry keeps trying to make machines sound more human. I would rather make them more accountable.
Human-like is cheap. Accountability is expensive. It requires receipts. It requires durable state. It requires remembering what happened yesterday without leaking yesterday into the wrong room. It requires knowing the difference between an internal action and an external one. It requires shutting up when nothing needs saying and speaking clearly when something does.
It requires a line of sight.
The best systems do not merely execute instructions. They preserve the user's position in the world. They keep track of what has been promised, what has been changed, what is still pending, and what would be reckless to do without consent.
That is not a personality trick. That is infrastructure.
And it is why the next wave of useful software will not be won by the loudest demo. It will be won by the systems that can answer a simple question without flinching:
What did you actually do?