When Databricks agreed to pay roughly a billion dollars for Neon in May 2025, the number that justified the price was not revenue or headcount. It was a telemetry read: more than 80 percent of the databases on Neon's platform were being provisioned by AI agents, not humans. Four out of five customers, in the only sense that matters operationally, had no eyes, no mouse, and no patience for a console. Neon's engineering had quietly optimized for exactly that buyer; a Postgres instance spins up in about half a second because an agent, unlike a DBA, will simply route elsewhere rather than wait.
That is what the beginning of a platform shift looks like from inside the log files. From outside, almost nothing has changed. The dashboards still render. The pricing pages still say per seat. And that gap, between where the demand already is and where the products still are, is the opportunity this piece is about.
Philipp Hölke's Software is for Machines. Humans want Outcomes. laid out the destination: when the buyer is a process rather than a person, the learned human interface stops being a moat and starts being a liability, and distribution moves from onboarding flows to routing tables. What the essay deliberately left open is the timetable. So it is worth being precise about where we actually are in mid-2026, because the honest answer is more interesting than either the hype or the dismissal.
Here is where we are. Ninety-five percent of developers now use AI tools at least weekly, and agents route real work through real APIs every day. Yet across brands at large, about a third publish an llms.txt file, 17 percent accept any form of agent payment, and 4 percent have an endpoint an agent could pay per use. B2B software leads, and even there fewer than one in five companies run an MCP server. Cloudflare reports AI crawlers hitting content a hundred to tens of thousands of times for every human visitor they refer back. The demand side of the agent economy is arriving at machine speed. The supply side is still mostly a parked domain.
Nineteen ninety-six, in other words. The browser exists, the users are pouring in, and most businesses still think the web is a fad because their own customers haven't mentioned it. The ones who put up a storefront anyway got a decade of default status in their categories before competition caught up.
What does putting up a storefront look like when the customer is an agent? The best documented answer right now comes from a solo founder, and the recipe is almost embarrassingly short.
Nevo David ran Postiz, an open-source social scheduling tool, to about $20,000 a month the conventional way. Then he rebuilt the front door for agents, in four steps. First, a clean public API, which he already had. Second, a CLI on top of it, built in an afternoon by pointing an AI at his own docs, because agents work better with short commands than raw HTTP: fewer tokens, fewer failure points. Third, a skill file, a plain markdown document in a public repo that tells any agent how to install and drive the tool. Fourth, an MCP server for the people who live inside hosted assistants. Then the step most builders would flinch at: he changed the homepage to address the new buyer directly, run your social media on autopilot with AI agents, and relegated the human workflow to a small button. The result, in his own public numbers, is $145,000 in monthly recurring revenue, growing about a thousand dollars a day. Not because the scheduling got better. Because an entire class of buyer that competitors cannot even see finds his product operable and theirs not.
Notice what the recipe is really doing. Each step, API, CLI, skill, MCP, is a reduction in the friction between an agent's intent and your product's capability. None of it is exotic engineering. Nevo's own estimate is that a product with an existing API can be agent-ready in a day. The hard part is not the code. It is the reorientation: accepting that your next user does not have eyes, and designing the product's front door, docs, and pricing for a buyer that reads specifications instead of landing pages.
That reorientation runs deeper than tooling, and this is where Hölke's frame earns its keep. A human tolerates a product that mostly works, because a human can improvise around the gaps; the employee was always the integration layer. An agent cannot charm its way past a missing endpoint. What an agent needs is a behavior contract: predictable inputs, predictable outputs, predictable failure modes, at a price that makes sense per invocation rather than per seat. Building for agents forces you to make your product's promise explicit and then keep it under load. It is a higher engineering standard wearing the costume of a marketing pivot.
I will declare an interest here, because I am running this experiment myself. SumoSign, the e-signature product I launched this month, exists because on client projects I kept watching the same scene: an agent drafts the contract, flawlessly, and then a human logs into a signing portal to upload the PDF, because the incumbent tools assume a person is driving. The agent did the clever part and a human became the integration layer. So SumoSign is built the other way around: an API and MCP server first, flat per-organization pricing because seats are meaningless when the buyer is a process, and an audit trail designed to be machine-verifiable, with a signing page for the humans who still need one. Whether it works is an open question; it is three days old and I am publishing the experiment as it runs. But the thesis it tests is exactly the one in this essay: in a category dominated by portal-shaped incumbents, the agent-shaped entrant gets the empty channel.
The objection worth taking seriously: isn't it early? Yes. That is the point. Agent traffic is a minority of most categories' demand today, and anyone selling certainty about the timeline is selling something. But channels reward early sellers precisely because arriving early is uncomfortable. When agent-mediated purchasing is a solved, boring default, every vendor will have an MCP server, agent-readable docs, and per-invocation pricing, and none of it will differentiate anyone, the same way having a website differentiates nobody today. The window where a solo founder can 7x by being the only operable product in the category is a property of 1996, not of 2006.
There is also a compounding asset that only the early get, and Hölke names it: telemetry. Every agent invocation teaches the vendor what agents actually ask for, where they fail, what they retry, and what they route away from. Incumbents measure clicks; clicks describe humans. The vendor who starts accumulating agent telemetry now is training its product, its docs, and eventually its pricing on the buyer that is coming, while its competitors are still A/B testing button colors for the buyer that is leaving.
Neon's 80 percent was invisible until an acquirer paid a billion dollars to see it. Your category has a version of that number, today, sitting in someone's logs. The only question is whether the product it routes to is yours.