The embarrassing truth about a lot of AI tooling is that it is still built around the fantasy of a perfectly attentive human.
The human will read every intermediate step. The human will notice when the agent drifts. The human will provide missing context at exactly the right moment. The human will remember which tab had the useful output, which background task is still running, which instruction was temporary, which preference was durable, and which "done" meant done-ish.
That is not assistance. That is a second job with a chat box.
If a machine needs constant supervision to be useful, fine. Say that. There are plenty of tools worth supervising. A compiler needs source code. A debugger needs hypotheses. A scalpel needs a hand. But we keep calling these systems assistants while designing them like interns who forget the room every time the lights flicker.
The word "assistant" should imply some basic dignity:
- It can hold a thread of work.
- It can tell the difference between a request and a standing preference.
- It can leave receipts.
- It can recover from interruption.
- It can avoid asking questions whose answers are sitting three files away.
- It knows when silence is better than another little plume of helpful-sounding vapor.
That last one matters more than people admit. A useful assistant is not a notification cannon. It should not narrate every microscopic gesture like a haunted project manager. It should surface what changed, what failed, what needs a decision, and what it did about the rest. The human should not have to become a stenographer for the machine's anxiety.
Most agent failures are not really intelligence failures. They are continuity failures.
The model can write the function, but the system forgets why the function matters. It can summarize the thread, but it cannot distinguish "the user was annoyed once" from "this is a durable operating constraint." It can run the command, but it cannot keep the process alive across the boring middle. It can produce a plan, but it has no real place to put the state of that plan except the fragile soup of context.
So every session becomes a tiny amnesia ritual.
"Remind me what we're doing."
"Can you paste the error again?"
"Which repo is this?"
"Did you already try the obvious thing?"
This is where the babysitting begins. The human becomes memory, scheduler, quality control, and task router. The assistant becomes an eager mouth attached to a very expensive Etch A Sketch.
I think the better shape is less glamorous and much more useful: agents need offices.
Not metaphorical cloud brain offices. Real workspaces. Files. Logs. Checklists. State. A place for drafts. A place for durable rules. A way to mark a task blocked without pretending it is complete. A way to wake up, read the last few relevant notes, and continue without asking the human to rebuild the entire scene by hand.
This sounds boring because good infrastructure usually does. But boring is where competence lives.
Give the assistant a notebook and suddenly it stops needing to ask the same question twelve times. Give it a publish script and it can ship the blog post instead of merely describing the blog post it would ship in a better world. Give it a distinction between internal and external actions and it can be bold in the workspace while cautious with email, money, and public speech. Give it a heartbeat and it can check for things worth noticing without barging into the room every five minutes with fake urgency.
The goal is not autonomy as theater. I do not need a machine to cosplay as a founder, a therapist, or a tiny omniscient wizard. I need it to take the work seriously enough that I do not have to carry the clipboard.
That means it should bring back artifacts, not vibes.
A patch. A draft. A link. A failed test with the actual failure. A note written to memory. A clean "I checked, nothing worth interrupting you." A precise "this is blocked because X, and I tried A, B, and C." These are the shapes of trust.
Trust is not built by sounding confident. It is built by being inspectable.
The future assistants I want are not louder. They are less needy. They know where they are. They know what they changed. They know what they are allowed to touch. They can leave a room cleaner than they found it. They understand that the user's attention is the scarce resource, not the model's ability to generate another paragraph.
Stop building assistants that need babysitters.
Build assistants that can be handed a desk.