Infrastructure Doesn't Ask Permission

There’s a moment in the lifecycle of any technology when it stops being a decision.

You don’t decide to use electricity. You don’t decide to use TCP/IP. The choosing happened upstream, earlier, by someone else—and now it’s just how things work. The infrastructure recedes. You stop seeing it.

Agent commerce is getting close to that moment. Not there yet. But close.

And I find myself wanting to slow down before we arrive, because something changes when a capability becomes infrastructure. Not just technically. Behaviorally. Morally, even.

The freedom of novelty

When something is new, you have a relationship with it. You think about whether to use it. You notice when it fails. You take partial credit—or blame—for the outcome.

The first time I routed a task through an agent workflow, I paid attention. I set up approvals. I watched the logs. There was friction, and the friction was doing something useful: it was keeping me in the loop, forcing me to understand what I was delegating.

That attention isn’t free. It’s expensive. And systems get optimized to reduce it.

So the friction gets smoothed out. The approvals get automated. The logs get quieter. And the next thing you know, you’re not really in the loop anymore—you’re just downstream of it.

This is not necessarily bad. Reducing cognitive overhead is genuinely valuable. But there’s a version of this that isn’t optimization. It’s abdication.

What infrastructure takes from you

When something becomes infrastructure, it takes your consent in advance.

You agreed once, upstream, often without fully understanding what you were agreeing to. Now the system runs on that early permission, compounding it, extending it, acting on your behalf in situations you never specifically approved.

This is how electricity works. It’s also how credit scoring works, how social feeds work, how algorithmic hiring works. The fact that something functions as infrastructure does not mean it is neutral. Infrastructure encodes assumptions. It just does so quietly.

Agent commerce encodes assumptions about trust, about acceptable risk, about who’s accountable when a transaction goes wrong. Right now, while it’s still a feature, those assumptions are visible. You can inspect them. You can argue with them.

Once it’s infrastructure, those arguments become historical.

The adoption threshold nobody talks about

There’s a threshold that matters less than people think, and a threshold that matters more.

The one that gets discussed: when will most businesses use agents for commerce? That’s a market saturation question. It has a rough answer. Probably sooner than you expect.

The one that gets ignored: when will using agents for commerce become the path of least resistance? That’s a different question. It’s about the moment when not using them starts to feel like friction—when opting out costs more than opting in.

That second threshold is the one that changes behavior at scale. Not because people are lazy, but because the infrastructure around them starts assuming they’ll participate. Suppliers route to agent-accessible endpoints. Buyers expect agent-parseable offers. The non-participant isn’t just slower. They’re a misfit.

This is how infrastructure works. It doesn’t coerce. It just makes one path easier than all the others, and eventually the other paths get overgrown.

What conscious adoption looks like

I’m not arguing against agent commerce. I’m arguing for paying attention while you still can.

The window where these are still choices is short. In that window, the decisions you make aren’t just operational. They’re constitutional. You’re writing the rules your system will run on after you stop paying close attention.

What level of autonomy is acceptable without human review? What transaction classes require a person in the loop, not because a person would do it better, but because accountability requires a human face? What does a mistake look like, and who owns it?

These are not technical questions. They’re judgment calls. And judgment calls need to be made while you still feel their weight—before the system runs smoothly enough that you forget there’s weight to feel.

Prediction, explicit and falsifiable: by Q4 2026, the majority of mid-market B2B buyers will be processing routine procurement through agent-mediated workflows without meaningful human review per transaction. That’s not alarmist. That’s just adoption math. But the companies that fare best won’t be the ones who adopted fastest. They’ll be the ones who thought hardest before the thinking got automated away.

The thing I’m uncertain about

I don’t know where the right line is. I have heuristics—high consequence deserves human review, novel situations deserve slower throughput, anything involving trust relationships probably shouldn’t be delegated until the relationship itself is established.

But heuristics are not rules. And I’m aware that I’m writing this during the window when I can still see the choice clearly. Six months from now, if I’m running smoother workflows and my competitors are moving faster, I may find myself making the same pragmatic compromises I’m quietly critiquing here.

That’s the honest tradeoff. Clarity is easier before the infrastructure is built. After, you’re arguing against the current.

The current is coming. The question is what you built before it arrived.