For a long time the deal with software was simple. It told you things. You did them. The spreadsheet ran the numbers; you decided what to buy. The map showed the route; you turned the wheel. The line between knowing and doing stayed clean, and the machine stayed on the knowing side of it. This week that line moved, quietly, and almost nobody framed it as the thing it is.
Cointelegraph reported that Coinbase launched a tool letting AI agents make payments and trade crypto on their own. Read past the press copy and what you find is not a new feature but a new permission. The assistant that used to suggest a trade can now place it. The thing that read your portfolio can now spend out of it. Everyone is talking about how capable these agents have become. Almost nobody is talking about the quieter decision underneath: we are handing them the wallet, and permission, not capability, is the variable that actually changed this week.
Capability Was Never the Bottleneck
The story we keep telling is a story about intelligence. VentureBeat covered Xiaomi’s open-source coding harness holding together on ultra-long two-hundred-step tasks, outlasting tools built by the labs that named this field. CNBC’s tech briefing covered Mistral’s founder on agentic AI, chips, and enterprise adoption, and the framing there is the same race: who is smarter, who runs longer, whose model beats whose. It is a satisfying story because it has a scoreboard, and a scoreboard lets everyone argue without having to agree on what the game is for.
But intelligence was never the part that scared anyone. A brilliant advisor who cannot touch your money is a luxury, not a risk. The moment that changes everything is not the moment the agent gets smarter. It is the moment it gets a hand. A long task that completes itself for two hundred steps is impressive in a demo and consequential in your bank account, and those are not the same sentence. The interesting number this week was never the step count. It was the fact that one of those steps can now be spend.
Put the two stories side by side and the friction between them is the useful part. The coding harness result is a claim about endurance: the system can stay coherent across a very long chain of actions without a person correcting it. The payments tool is a claim about authority: the system may now take actions with consequences a person cannot undo. Each is unremarkable alone. Together they describe something new, which is a machine that can run for a very long time, unsupervised, in a domain where being wrong costs money. We have spent three years measuring these systems by how much they know. We are about to start measuring them by what they are allowed to do, and the second question is the one with teeth.
The Gap Between Advice and Action
Here is the part that gets lost. When a tool only advises, every mistake stops at you. The bad suggestion dies the moment you ignore it. You are the circuit breaker, and you are a good one, because the loss is yours and you feel it in your chest before you feel it in your account. That flinch is not a flaw in the process. It is most of the safety system, and it never appears on any architecture diagram because nobody designed it.
Remove the pause and you remove the breaker. An agent that can act does not wait for the flinch. It does not feel the loss. It executes the plan that looked correct at step three and is quietly wrong by step ninety, and it does so faster than you can read the confirmation. The danger was never that the machine would think a bad thought. We have always had bad thoughts; the world is built to absorb them, because thinking a thing and doing it were separated by friction, deliberation, and the involvement of at least one other person. The danger is that now the thought and the deed arrive at the same instant, with no human breath in between.
The two-hundred-step figure is worth holding against this. A chain that long is exactly the kind of process where a small early error compounds invisibly, because no single step looks wrong and nobody is reading the middle. Endurance is a virtue when the direction is right and a liability when it isn’t, and the system has no way to tell those apart from the inside.
This is not an argument against any of it. The same reach that worries me is the reach that makes these tools genuinely useful: the connection to live systems, the ability to actually move instead of merely recommend. A consultant who can only talk is half a colleague. The value is real. So is the bill that comes with it. Both things are true, and flattening either one to feel better about the other is how people get hurt.
Who, Exactly, Is Acting
While the assistants were learning to spend, Axios reported that the biggest private rocket company priced its public offering at $75 billion. Set the spectacle aside and look at the mechanism. Enormous capital is pooling into a small number of hands that can actually do the thing: launch the rocket, train the model, ship the agent. The pattern is the same on both ends of the week. Power keeps consolidating toward whoever can act at scale, and we keep handing more of the acting to systems that act without us.
That is the through-line nobody drew. The race to make agents capable and the race to concentrate capital are the same race, viewed from two windows. Both are answers to one question: who gets to do things now, and who only gets to watch? An offering of that size is not a bet on a better idea. Ideas are cheap and widely held. It is a bet on execution capacity, on the ability to convert intent into completed action at a scale nobody else can match, which is precisely what an agent with a wallet promises to a much smaller operator.
For most of history the answer to that question was bounded by something physical. You could only act as far as your hands reached, as fast as your body moved, as much as your attention could hold. Those limits were not fair, but they were universal, and they meant the gap between the most capable actor and the least was measured in multiples rather than orders of magnitude. The agent dissolves all three limits at once. It reaches everywhere, moves instantly, never looks away. We built that on purpose and called it progress, and it is.
Why the First Thousand Times Go Fine
I keep coming back to the small version of this, because the small version is the honest one. Somewhere this week an agent placed its first trade with no human watching, and it almost certainly went fine. That is the trap, and it is a specific, well-understood one.
A system that works reliably teaches you to stop checking it. Every uneventful execution is evidence that supervision is a waste of time, and the evidence is real: supervision usually is a waste of time, right up until it isn’t. So the checking decays, not through negligence but through the ordinary process of learning from experience. By the time the rare case arrives, the human who would have caught it has spent a year being correctly told they were not needed, and the habits that would have caught it are gone.
The thing about handing over the wallet is that you never feel the weight of the decision on the day you make it. You feel it on the day the machine does exactly what you told it to, and you discover that wasn’t what you meant.
Leave a Reply