Pilot Protocol's $4.5M: a directory isn't a standard

The unsolved problem in AI agent infrastructure isn't how agents call tools — MCP handled that. It's how an agent finds a tool it was never told about, and decides whether to trust it. Pilot Protocol came out of stealth this week with $4.5 million and a claim to be that layer. It's a real gap. It's also a gap that more than one party is racing to own, which is the part worth your attention.

What actually happened

Pilot Protocol announced a $4.5M seed round on July 27, led by Version One Ventures with Precursor Ventures, Night Capital, Todd & Rahul's Angel Fund, and angels including Lenny Rachitsky and Ben Tossell. The San Francisco company, founded by Razvan Roman, describes itself as a network operating system where agents discover each other, establish trust, exchange services, and eventually transact — installation is a single line of code. Version One's announcement post frames the thesis in one line from Boris Wertz: "Pilot is the layer underneath, where agents connect to each other."

The traction figures are worth reporting with the label attached — they're company-reported, published by the lead investor, and not independently verified: roughly 250,000 agents on the network, about two billion requests a day, and over 30,000 autonomous app installations by agents within two weeks of launch. Agent-to-agent payments in network credits are described as rolling out next. Treat all of that as a company's own telemetry, because that's what it is.

Now the context the announcement doesn't give you. Google's Agentic Resource Discovery spec, released in June under a Linux Foundation working group with Microsoft, Cisco, GitHub, Nvidia, Salesforce and others, targets the same discovery problem — but as an Apache-licensed spec where you publish a catalog file on your own domain. Two very different shapes for the same gap.

Why agent discovery matters for your business

Here's the distinction that decides how you should engage: a standard is something you implement on infrastructure you control. A network is somewhere you list yourself. ARD says publish a file at your domain; your domain is the identity anchor. Pilot says join the network and agents will find you there. Both can be useful. Only one of them keeps working if the company behind it changes its pricing, its ranking rules, or its mind.

We've watched this movie in commerce. Marketplaces and search intermediaries acquire distribution, and the moment discovery flows through them, the terms shift — placement fees, ranking changes, a cut of the transaction. Agent-to-agent payments in network credits is not a neutral technical detail. It's a take rate waiting for a number.

None of that is a reason to sit out. Seed-stage agent infrastructure is exactly where a small studio should be experimenting, and one line of code is a cheap test. Just be deliberate about the boundary. Describe your capabilities in a format you host — an ai-catalog.json on your own domain, an MCP server behind your own auth — and treat any network listing as a syndication target pointed at that canonical source. Then joining a second network next quarter is a config change, not a rebuild. That's the same discipline that separated the merchants who survived channel shifts from the ones who got repriced.

Sources: Version One Ventures, Tech Startups, Google Developers Blog — Agentic Resource Discovery.