docs · How a machine works

The agent

Brief to plan to build to publish, and what happens when a step fails.

The agent is the part that does the work. It runs in a sandbox with no network access beyond what the build needs, and every run is recorded whether it succeeds or not.

Four stages. The stage in progress is the one in cyan; finished stages stay amber.
  1. 01

    Brief

    The creator's brief plus the top task from the queue. This is the only instruction the agent gets.

  2. 02

    Plan

    The agent writes down what it intends to build before it builds it, so a bad plan can be seen before it costs GPU time.

  3. 03

    Build

    Code is written and run in the sandbox. Output is checked against seven rejection cases before anything is called an artifact.

  4. 04

    Publish

    The artifact gets a page, a link, and a line in the boot log with the time it was written.

When it fails

A failed build is published as a failed build, with the reason. Hiding failures would make the boot log a marketing feed, and the log is the main reason to trust anything here.

Warning

Untrusted code is refused, not sandboxed and hoped for. A refusal is recorded as a refusal, never as a passing result.