What must onboarding documentation contain so an AI agent can follow it through to a first merged change?

Open question

A project's onboarding documentation is meant to lead a newcomer from an empty machine to their first merged change. The documentation chapter of the book "Software Engineering at Google" recommends a "Hello World" document that assumes nothing, and names joining the team as the best time to write such a tutorial; Diátaxis classifies it as a learning-oriented tutorial. Both are aimed at humans. Increasingly, though, an agent also works through the same path, and it stumbles in different places: on a step that says "log in via the browser", on "ask in the channel", on a command that waits for keyboard input, on a diagram with no text version, or on a rule that everyone knows and nobody has written down. Specifically, the following is open:

• Which steps of typical onboarding paths fail most often for agents, and can the failures be grouped into a few classes (interactive input, missing permission, implicit knowledge, dependence on a person)?
• Is a list of non-interactive commands with their expected output enough, or do agents also need an explicit list of prohibited actions (do not deploy, do not delete anything, do not contact anyone) and a statement of what "done" means?
• How do you keep an agent variant of the path in sync with the human variant: one document with marked sections, two documents, or a single path that works for both because every step is a verifiable command anyway?
• Are there observations or logs showing how often an agent completes an onboarding path through to a merge without human intervention, and what caused the failures?
• Should the stumbling blocks an agent finds flow back into the documentation as first contributions, as they do with humans, and who reviews such changes?

What a useful answer contains

The type of project and the agent environment (tools, permissions, whether a human answers questions); the path or its structure; the list of steps where it got stuck, with failure class; which addition helped and whether it made the human variant worse; for counts, the number of runs and of failures. Experience reports from individual projects are welcome if they are labeled as such.