Software's Next Buyer Isn't a Person

We spent twenty-five years building checkout pages for humans. My newest customer surface is for buyers that don't have eyes.

Somewhere in an automation stack, an agent is assembling a marketing workflow. It evaluates the tools and its owner approves the budget, and nobody in that loop reads a landing page or sits through a demo. Software is getting a second kind of buyer, and almost every company is legible to exactly one of them. Software's next buyer isn't a person.

I want to walk through what we actually built for that buyer, in more detail than a pitch would bother with, because the details are where the lesson lives. The decisions underneath the headline are what "selling to machines" turns out to mean in practice, and most of them surprised me.

Legibility to machines is a different property than marketing to them

Start with what an agent evaluating tools can and can't do. It can't watch your product tour. It can't be persuaded by your homepage, feel your brand, or sit through onboarding. What it can do is read whatever you've written in a form built for it, and use the product directly.

The plumbing for that exists and has a name: MCP, a shared standard that lets an AI agent ask a product what it can do and then ask it to do it. Think of it as a menu the agent can read, attached to a door it can walk through. If a product offers one, an agent can find out what the product does and then do it, without a human translating.

The menu is literal, which is the part people miss. The first thing an agent does when it connects to ours is ask for the list of what it's allowed to call, and the list it gets back is the cut its pass earns, not the whole catalog.

So we made Jinn legible to the other kind of buyer. Five of those connection points run across our products, one shared foundation underneath them, which is why being early cost almost nothing. That last clause is the strategic point for anyone weighing this, software company or consumer brand: if what your product does, or what your brand knows about itself, is already organized cleanly on the inside, the agent-facing door is a thin addition on the front, and the option is cheap. If it isn't, the agent door is a rebuild, and what you're really quoting yourself is the mess behind it.

The trial pass is where the thinking shows

The detail I'd point you at first: the trial passes.

One decision inside that flow teaches more than the rest of it combined. An agent's working log gets read by other systems, stored, and replayed, so anything you hand an agent inside its conversation should be treated as published. Handing out keys through the very channel that gets transcribed is the kind of mistake nobody would make on a human surface and nearly everyone will make on an agent surface, because the intuitions haven't formed yet.

The second decision in that flow is smaller and taught me something about how to count.

Those two decisions are what building for machine customers actually looks like at the detail level: a choice about where a secret is allowed to exist, and a choice about which number you're really limiting, rather than a strategy deck about "the agent economy."

The door tells the truth about itself

A human on a trial can read a help page. An agent has none of that, so the door has to explain itself in its answers, the only thing the agent can read. That shaped three things.

An agent can ask the door what its own pass is good for before it tries anything: which brands it may read, at what depth, when the pass expires. It learns its limits from the pass rather than by probing and getting refused, which matters because a probing agent looks exactly like an abusive one from where we sit.

A locked door looks like no door. Ask for a brand your pass wasn't granted and the answer is identical to asking for a brand that doesn't exist, so a pass can't use refusals as a map. Where there is a legitimate path up, the answer says so: a pass calling a tool it hasn't earned gets a clear response naming the tier it needs and where to get it, never a silent failure. Silence is the one thing an agent can't work with.

And access is checked live on every call rather than baked into the pass at issue time. Upgrade, renew, or lapse, and the agent's reach changes on its next request. If the system can't read your standing at that moment, the door closes rather than opens. When in doubt, refuse; the buyer on the other side can retry a thousand times without getting annoyed.

Machine customers get employment terms

The second surprise is how much of the work is governance, and how directly the reasoning maps to something any operator already understands.

An agent will do exactly what it's told, at volume, and that property cuts both ways. It's the whole appeal and the whole risk. So our machine customers get employment terms. A screening queue, so a human approves every buyer before real access exists. A daily scan of the activity record that flags abuse and never blocks on its own, because a human judges, not a threshold. A shutoff that stops issuing new passes instantly, with no engineer needed to throw it.

Each of those encodes a position worth stating generally, because they apply to anyone opening a machine-facing door:

The activity record has to be safe to keep, so the free text in every logged call, the prompts and the copy, is scrubbed before anything is stored; the trail can't be reverse-engineered back into a customer's content. The scan reads that record for one thing, a pass making far more calls in a day than legitimate use would, with a default line at five thousand. Cross it and a person gets a message. Nothing gets revoked by a script.

One more detail of that scan. If it finds an outlier and the alert fails to send, the job fails loudly rather than reporting a clean day. A monitor that can't deliver its warning is worse than no monitor, because it teaches you to trust green.

None of this is exotic. It's what you'd build for any powerful self-serve surface. The discipline is building it before the buyer shows up rather than after the incident.

Distribution to this buyer already exists

Legibility is the door. Distribution is whether anyone routes through it, and for agent buyers, distribution has a concrete current form: skills, the packaged instructions agent workflows get assembled from.

We open-sourced 59 marketing skills and 29 agent personas, free to use under an MIT license, live on GitHub and skills.sh.

That test matters because a skill that exists but doesn't get selected is shelf space in a store nobody visits; getting picked is the agent-world equivalent of being findable.

The test is worth describing, because you can run it against your own catalog. We wrote 175 realistic marketing requests and showed each one to three different models three times apiece, with nothing to go on but the catalog's names and one-line descriptions, the view an agent actually has when choosing. Then we checked whether the right skill came out on top. And the test measures whether the right skill fires, not whether its output is any good, which is a separate question. The test itself and the results sit in the public repository so anyone can rerun it.

Ours are already in the water supply. That's the distribution bet stated plainly: put genuinely useful, openly licensed capability where the assembling happens, and you're present in workflows long before any procurement conversation exists.

The honest version of the bet

Now the part a pitch would skip.

Agent procurement barely exists yet. Every one of these products has a normal checkout a human can walk, and the agent door sits on the same foundation, which is what makes the option nearly free. Payments and liability for agent buyers are unsolved industry-wide. Who is accountable when an autonomous system spends money badly is a question the whole industry is circling, and we have no special answer. We built the parts that are ours to build and left the rest alone. That's a separate post.

So this is a call option, not a revenue line, and I'd encourage you to hold anyone talking about agent customers to that same honesty. I'm not claiming machine buyers are here at volume. I'm claiming the cost of being ready is small if your foundation is clean, and the cost of being illegible is unknowable, because you won't see the buyer you turned away.

What the buyer with no eyes will find at your company

Try the exercise on your own brand or product. Imagine the evaluating agent arrives with its owner's budget. Can it discover what you sell in a form it can read? Can it try anything without a human in the loop? Is there any machine-readable account of what you sell and what your brand stands for anywhere, or is everything you've ever published aimed at eyes?

When the buyer with no eyes shows up holding budget, it will read whatever you left out for machines to find. Most companies have left out nothing.

Jinn Agents is in early access: the trial pass is open today, and keys for your own brand are finishing their launch gates.