Search for the artifact

I needed to find companies that publish specific accounts of their own AI engineering work. My agents run the searches under my direction: I set the question and the standard an answer has to meet, they search and screen, and I judge what comes back. What improved here was the instruction they worked from.

The first search described the subject. It combined AI agents, internal systems, production, cost, and postmortems. Seven results came back. Every one was a vendor cost guide. The titles promised a full breakdown, a complete budget, or the cost of AI agent development in 2026.

There were zero company engineering blogs and zero first-hand accounts.

The search had done what I asked. Those are the words a buyer uses, so vendors build pages around them. I had described a market topic and received the pages competing for that market.

The second pass described the artifact I wanted to read instead of its subject: phrases such as “we built” and “we shipped,” a named internal system, a measured result, an incident account, or a multi-part engineering series. The subject stayed the same. The results changed kind.

Stripe’s engineering blog returned a series about Minions, its internal coding-agent system. The second part describes “blueprints,” workflows that combine deterministic code nodes with agent loops that can make their own decisions.

It also gives a large figure. More than 1,300 Stripe pull requests merged each week are produced by Minions and contain no human-written code.

The secondhand summary stopped there. The verification pass my agents run against the primary source restored the condition that matters: every one of those changes is human-reviewed. The pull request contains no human-written code and still carries human judgment. Repeating the number without its review condition would turn a precise operational claim into a broader claim the source never made.

Shopify’s engineering blog returned Under the River. The post names three internal systems. World is its monorepo. Aquifer is its platform for running agents. River is an agent built on that platform.

The post reports that River coauthored one in eight merged pull requests across Shopify. In a recent 30-day period, it ran 59,918 sessions. Those figures sit beside an explanation of what the company had to build for the system to work.

Shopify writes, “Every change we made for agents was also the right thing for humans.” It names the monorepo, reproducible environments, written skills, and clean CI feedback. Those improvements had already been wanted and had not been urgent enough to receive resources.

Stripe describes the same sequence in its own system. Minions use the standard remote development environment already used by Stripe engineers. The environment was designed to be parallel, predictable, and isolated for human work. Agents inherited those properties.

These pages appeared because the query asked for evidence produced by practitioners. Named systems, implementation details, incident language, and figures with conditions are expensive for a seller to imitate consistently. They are ordinary features of an engineering team describing work it completed.

The source also matters. An investor or vendor writing about a company proves that the company reached an investor or vendor. It does not prove that the company’s engineers explain their work publicly. For this search, the signal was an engineer publishing under the company’s own name about a system the company built.

What the searches find gets graded in public. The AI-first companies registry lists no company today. What is published so far is the bar a company has to clear: six dimensions, per-claim evidence a stranger can open, admission at 7 of 12. Grading is a person reading that evidence, thirty to sixty minutes per company, and the first candidates are queued for it. The page is where results will land.

The method applies beyond company blogs. A broad debugging search describes the subsystem and returns popular explanations. Searching the exact error, the operation that triggered it, and the measured condition around it is more likely to find an incident record. A prior-art search improves when it asks for a named implementation, a migration account, or a result with a constraint. Searching whether a problem has been solved improves when the query resembles the record a solver would leave.

So when a search comes back full of sellers, the instruction I give is no longer another topic word. It is a description of the artifact I expect a practitioner to have written.