Agents are the primary customer

Watch where the fastest-growing infrastructure companies are getting their revenue and you’ll notice something odd: they aren’t acquiring users. They’re acquiring agents.

AgentMail sells email sending to AI agents. Composio sells three thousand integrations behind one API key and raised $25M. Resend, BrowserBase, Exa - the same pattern. A developer wires the product into an agent once, and from that moment the agent consumes on its own schedule, with nobody to re-convince.

I said this out loud in a meeting a few weeks back, half thinking aloud: it’s much easier to get agents as your customer, because if you acquire actual humans, it will take forever. The more I’ve built since, the less it feels like a quip.

Strictly, the customer of record is still a human or a company - someone pays the bill and carries the liability. But the agent is the one reading your docs, making the calls, and deciding whether to keep calling. When I say customer, I mean the one you have to win.

The economics underneath

Human acquisition is a war of attrition. You pay for attention, you fight for a habit, you lose most of the funnel, and the users you win can quietly leave any week they like. Every SaaS metric we’ve built for twenty years - CAC, activation, retention curves - is really a measurement of how hard it is to change one person’s behavior.

Agent acquisition collapses all of that into a single event: a developer makes one configuration decision. Add the API key, grant the scope, write the line of config. There’s no habit to build, because agents don’t have habits - they have schedules. There’s no churn from boredom or forgetting, because software doesn’t get bored. Churn still exists - a key gets revoked, a benchmark starts favoring someone else, a model update reroutes every agent at once. What disappears is the need to re-persuade anyone.

I’ve felt this from the buying side. A gateway service I wired into my own agent stack at its free tier - within weeks my agents were using multiples of the intended limit, not because I decided anything, but because agents run whether or not I’m paying attention. When that product starts charging, I’m not a lead to be nurtured. I’m revenue that already happened.

What it changes if you’re building

If the agent is the customer, the product decisions invert. Documentation matters more than a landing page, because the “buyer” evaluating you might be a coding agent reading your docs. Onboarding has to survive being done by software: one key, one call, a response clean enough to parse. Pricing wants to follow usage, because usage is what compounds - a seat is a human concept.

And distribution stops meaning attention. It means being the default answer when an agent asks “what do I use for X.” That slot compounds while you hold it, and it can vanish fleet-wide in one model update - which is why services have started benchmarking themselves. Two weeks after this essay went up, Supabase shipped Evals, a public score for how well coding agents build with Supabase. Paul Graham’s read: one day every service used by agents will do this, “since those that can’t be used by agents will go out of business.”

The honest limit

This flywheel only spins where the underlying system lets software in. The conquered categories - email, search, crawling, integrations - were conquerable because APIs existed or could be built without anyone’s permission. The unconquered ones, like real commerce execution, stay unconquered because the platforms that own the data won’t let agents through the front door. Stripe and OpenAI tried - Instant Checkout, September 2025 - and by March OpenAI had moved ChatGPT shopping back to product discovery, with fewer than thirty of Shopify’s millions of merchants ever live. Google is making the next attempt with its own protocol. That’s not a footnote; it’s a map. The hard, closed categories are exactly where the next Composio-sized outcomes are hiding, priced accordingly.

A human buys once, after a quarter of persuasion. An agent buys every hour, forever, after one line of config. Build for the second customer.

⁂