Skip to content
Partifact

Overview

On 7 June 2026, a coding agent was handed an empty folder, a one-paragraph request, and a documentation URL. Thirty-nine minutes later it delivered a working parts-ordering app. It provisioned its own credentials, placed and confirmed an order with a supplier, and verified that every live status update genuinely came from the platform. No human wrote a line of code.

Partifact is the small experiment behind that run. It started from one question:

What changes for an API when its consumer is an AI agent?

Agents now arrive on both sides of an API. Some write the code that integrates with it. Some use it directly to get work done. Partifact looks at one narrow slice of that: same API, same task, two documentation surfaces, and an honest count of what differed.

Two kinds of agent

Section titled “Two kinds of agent”

Two words carry this site:

  • The builder creates an integration: a client, a webhook receiver, a UI. It works from the documentation alone.
  • The operator uses the API to do business work: open a repair job, order parts, settle the invoice. It calls tools and never writes code.

Each got its own test.

The builder test

Section titled “The builder test”

A coding agent in an empty folder was asked to build a “Parts Desk” web app for a repair shop, with nothing but a documentation URL. The test ran twice: identical prompt, identical API, identical model. The only variable was the documentation. One run got human-style HTML pages. The other got the same content as machine-readable files and searchable doc tools.

Human-style docsAgent-native docs
Working, security-verified appYesYes
Active build time~58 min~39 min
Messages exchanged394186
Tool calls16583

One run per configuration, against a deliberately compact API: one workflow, about a dozen methods. The numbers are directional and scoped to that. The full set, with every caveat: The numbers.

What stood out:

  1. Both runs delivered the app; the documentation changed what delivery cost. Half the conversation, half the tool calls, a third of the errors.
  2. The relationship to documentation changed in kind. The HTML-only agent scraped every page and reconstructed the contract. The agent-native one asked it questions.
  3. The agents surfaced three real bugs in the API that earlier testing had missed. All three are documented with the results.

Full story: The builder test.

The operator test

Section titled “The operator test”

The same API, used rather than built against. Given only the procurement tools and their descriptions, an agent took a collision-repair job from damaged vehicle to reconciled invoice with no human help, including recovering from a deliberately induced order error on its own.

Full story: The operator test.

What this is

Section titled “What this is”

Partifact is a small, self-contained demonstration built in a few days: a realistic sandbox of an automotive parts-procurement API, a typed SDK, machine-readable docs, and two MCP servers. It is a working sketch, not a product. What is faithful to the real contract and what is simplified is listed on Real vs. simulated.

Choose your path

Section titled “Choose your path”