I Trust My Car More Than My AI Agent. That Gap Is Where We're Going.

TLDR: A developer who runs his own AI agent daily finds it makes him more productive and more watchful at the same time. The gap between trusting a car and trusting an agent is real, explainable, and not permanent.

Summary: The writer built and deployed his own AI agent on a dedicated machine. It does real work every day. And yet he watches it more carefully than he watches his car. That tension is the starting point for an honest look at where agent technology is right now and where it is going.

The trust comparison is the part that stuck with me. When you get in your car, you do not consciously decide to trust it. You drove 100,000 km. The engine started every time. You learned the sounds it makes on a cold morning. Trust accumulated without you noticing. With a deterministic agent doing the same boring file rename every night, that same accumulation happens. Boring is exactly the point. But when the task is open-ended, involves judgment calls, and requires recovering from its own mistakes, the math changes. Outcomes spread. The black box gets blacker. Trust does not build the same way when you cannot predict what the box will do next.

The current technical ceiling is not a secret: context limits, memory that forgets things it should hold, retrieval that grabs the wrong document at the wrong moment, tool calls that fail without telling you, an agent that gets stuck without knowing it is stuck. None of those are permanent problems. They have been improving in jumps, not on a clean schedule, but improving. The boring future, steady rather than discontinuous improvement, is actually the one that changes how most people live. The frontier gets the headlines. The floor is what moves lives.

Right now agents are still a nerd hobby. Most people who say they use AI mean they typed into a chat window. Most companies claiming to be AI-first are in the large majority that have almost nothing real deployed. The on-ramp is the missing piece, and that is exactly what Apple's WWDC overhaul is betting on. Multi-step task chaining wired into the App Store, built on Google's Gemini underneath, which is a signal that the raw model is becoming a commodity. The product is the assistant layer on top. When that lands on hundreds of millions of devices, the open web stops being something humans browse. Agents do it, at machine speed, and the traffic numbers will look very strange.

The infrastructure cost reality is not soft-pedaled here: inference is already about two thirds of all AI compute this year, up from a third in 2023. Data center electricity demand is climbing double digits annually. GPU prices are not dropping. The author keeps a local model on a cheap Mac mini so he has a floor that does not move when the market does. That is a reasonable hedge. On the jobs side, the World Economic Forum projects 170 million new roles and 92 million gone by 2030. Net positive. Cold comfort if you are one of the 92 million. The honest answer is that this is a transition that asks people to move between kinds of work, and people do not all move at the same speed. Policy matters here. Models do not fix that. On the physical side, home robots are arriving, from 1X's NEO to Tesla's Optimus on BMW's factory floor. Useful home robots by 2028 to 2032. Close enough to think about now, far enough that you should not panic today.

Key takeaways:

Trust in AI agents builds fast for deterministic tasks and slowly for open-ended ones, because predictability is what trust actually requires
The raw model is becoming a commodity; the assistant layer on top is where the product differentiation lives
Inference costs, energy demand, and the jobs churn are real and happening now, not hypothetical future concerns

Why do I care: As someone who builds and ships software, the trust gap is the most practically useful frame I have seen for explaining why agent adoption is slower than the demos suggest. Demos show the best case. Ops shows the black box failing quietly at 2am. The author's point about getting good at directing this stuff rather than racing it is where I land too. The engineers who figure out how to point agents reliably, catch the quiet failures, and build the on-ramp for normal users are going to be more valuable than the ones optimizing model weights. That is a career framing worth holding onto.

I Trust My Car More Than My AI Agent. That Gap Is Where We're Going.