The Agent in Your Pocket Is Almost Useful
Personal AI assistants can finally do things, not just say things. A field report from the messy, promising gap between demo and daily driver.

Somewhere in the last couple of years, the assistant on your phone crossed a threshold: it stopped being a search box with manners and started being able to actually complete tasks — book the thing, move the meeting, chase the refund. The demos are dazzling. Daily life with one is more complicated, and more interesting.
The honest scorecard looks like this. For tasks with clear inputs and forgiving failure modes, the agent is genuinely good — the kind of good where you stop noticing it. For anything involving ambiguity, multiple services that refuse to talk to each other, or consequences you'd hate to explain later, you become a supervisor of a very fast, very confident intern.
The supervision tax
That supervision is the hidden cost the demos never show. An assistant that's right 95 percent of the time still needs you to know which 5 percent to catch — and knowing requires roughly the attention the task would have taken. The breakthrough isn't raw capability anymore; it's calibration. The agents that feel trustworthy are the ones that know what they don't know and say so before acting, not after.
The best assistant isn't the one that never asks questions. It's the one that asks exactly the right one.
The plumbing is the other honest problem. The world's services were built for human hands, and every workaround an agent uses to click through them is a small treaty that can be revoked by a redesign. The real infrastructure shift — services exposing themselves to software agents on purpose — is happening, but in the slow, grudging way infrastructure always does.
Why the gap is the good news
Here's the optimistic read: "almost useful" is the most valuable stage a technology passes through. It's where the failure modes get discovered by early adopters instead of everyone, where the norms get argued out, and where the plumbing gets built under pressure from people who actually want the thing.
The pocket agent will cross the gap — the trajectory is unambiguous. The only question worth debating is what we'll wish we'd standardized while it was still on this side.