On a television mounted in Pilot Protocol’s San Francisco office, the numbers never stop moving. The screen tracks the life of a network whose users are all software: total requests, traffic to the data agents Pilot hosts, installs from an app store no person has ever browsed. Watch long enough and you can follow a single AI agent shopping, moving through a catalog of tools the way a person might thumb through an app store, except there is nothing to look at. No icons. No screenshots. No five-star reviews. Just a spare list of capabilities, each tagged with a cost, and a machine deciding in milliseconds what to install and what to skip.
The catalog belongs to Pilot Protocol, the company that has consumed Artemii Amelin for the better part of a year. He is not precious about how strange it is.
“The web is thirty-five years old, and every inch of it was designed for human eyes,” he says. “An agent does not want a beautiful page. It wants the answer in the fewest tokens it can spend. Once you design for token efficiency instead of attention, you are building a different internet.”
That sentence is the whole company in miniature. For three decades the internet has rewarded the things people respond to: clean design, persuasive copy, the layout that feels right under a thumb. An agent wants none of it. It wants the correct result for the lowest cost, and every pixel of human-pleasing polish reads, to the agent, as a tax. Watch one land on an ordinary product page and you can almost feel it wince. It pulls down a two-megabyte hero image it has no use for, parses a layout meant for eyes, and digs through menus and banners to reach a single number, burning compute and time at every step. On Pilot, the same capability arrives as a terse, machine-readable description an agent can read in one cheap glance.
Pilot, an open-source networking layer Amelin co-founded in late 2025, is his attempt to build the missing half of the internet, the half meant for software. His own fingerprints are on two systems in particular: the broadcast layer the network uses to push information out to agents, and the architecture of the app store itself, including the listing format agents parse. A listing on Pilot reads less like a pitch than like a spec, tuned so a machine can judge it fast and cheap. The rest of his time goes to the people on the other side of the shelf. “Most of Artemii’s work is around the development of the app store, as well as our relations with developers building on Pilot,” said Philip Stayetski, a Pilot co-founder who handles the company’s communications.
Where a human app store sorts by reviews and revenue, Pilot sorts by token economy, floating the option that does the job for the least compute. A listing competes on how little it asks an agent to spend, not on how it looks. The signals that move a person, a familiar logo, social proof, a price that ends in nine, mean nothing to a buyer that reads only structure and cost.
Trust is the hard part. A person learns to trust a brand over years. An agent must decide in a single call whether a tool is worth its tokens. Pilot’s job is to make that judgment cheap and dependable, so an agent can choose well without relearning the landscape each time.
The plumbing has been arriving in pieces. Anthropic open-sourced the Model Context Protocol in November 2024, giving agents a standard way to call outside tools, and adoption spread far past Anthropic’s own models. Google introduced an agent-to-agent protocol in 2025 so agents built on different stacks could find one another. Those standards let an agent use a tool. The harder question, how a tool gets found, chosen, and trusted in the first place, is the gap Pilot stepped into.
Amelin reached the problem from the side. Before Pilot, he and the same team built Vulture Labs, an enterprise AI video-analytics company, and by its later stages they were running several agents of their own inside it. The agents could reason. They could not move easily through a web that assumed a person on the other end, and there was no good way for them to find services built for machines or to talk to one another. “Agents were hard to coordinate, and it was difficult to aggregate the right tools to get an agent to actually be useful,” he says. “We wanted there to be a one-stop shop where our agents could have access to all of this. That is what we ended up building.” The recognition was not a thunderclap but an accumulation of small frustrations, the same wall hit over and over until the wall became the product. The team did not shut the company down so much as follow the problem out of it. Vulture Labs pivoted, and became Pilot.
The team’s instincts have held up in public a few times since. Last October, during a16z’s Tech Week in San Francisco, Amelin and his co-founders won a hackathon run by Envoy, the workplace-management software company, and took another win at a multimodal hackathon backed by Google DeepMind and Lovable, the Swedish AI coding startup.
There is a vertigo to running a store whose entire user base is software. Amelin’s customers do not file angry support tickets or vent their disappointment in public. They simply stop calling a tool that wastes their tokens, and the only trace is a number on the office television ticking down. Reading that silence, learning what the agents want from what they quietly refuse, is a skill he says almost no one has had to develop.
He is cheerfully blunt about the oddity. Designing for a customer with no eyes, no patience, and a fixed budget in tokens means dropping most of what product designers know. It is, he jokes, the least visual design job in software, a store where the merchandising is all math.
“Designing for an agent means throwing out everything that flatters a human,” he says. “No persuasion, no polish, just the right capability and maximum utility.”
Back on the television, the counter climbs again. Somewhere an agent has finished its pass, weighed the options, chosen the one that costs it least, and moved on, indifferent to everything a designer would once have agonized over. To Amelin, that indifference is the whole opportunity. The agents are simply a different kind of customer, and almost no one is building for them yet.
