The article discusses whether internal documentation should be written for human readability or optimized for AI agent efficiency, exploring trade-offs and advocating for context-specific formatting.
I keep documentation for things like runbooks, ADRs, lab-notes, project specific documentation. I have files written on what was done in a session, what broke, what worked, what changed, what lessons were learned. My agents do the majority of the writing, and also maintain the documentation. We retire some documentation when its appropriate to do so, sometimes its archived, sometimes its superseded by a new document which contains the still-relevant parts from the retired document. Human Facing Documents: These documents require formatting that make things easier for me, the human, to understand. If something needs to be written that I will want to read and possibly comment on, then I want it written in a way thats most accessible for me. Things like definitions, the proper amount of context so I dont need to go searching, specific formatting requirements for easy readability. Some, specific, data and values right in the document where relevant. The bottom like is that I want human-facing text to be written in a way that will allow me to understand what i'm looking at, including the surrounding context that I would need to properly understand the purpose and meaning of the document. Agent Facing Documents: Agents dont need what I need. I use google's OKF and they can, more easily than I, pull in the additional files that are linked and grab what they need. They have the time to read though glossary files and index files and collect everything together for themselves. They dont need the context-bloating additional nice-to-haves that I want for my own purposes. If every document was written for me, then that'd be less efficient for documents that I'm unlikely to ever read. The question here is, do I have the documentation written with the type of reader in mind or should all documentation be written with a human in mind? The distinction being that documents written with an agent in mind can be more token efficient than one that needs to be easy for a human to read. With the understanding that there will be some documentation written thats unlikely to have human eyes ever look at it, but parsed by many agents.
The article argues that designing APIs for AI agents requires different principles than for humans, emphasizing clarity, explicitness, and avoiding defaults, because agents can read entire docs and write extensive code.
Jerry Liu discusses the need for an agent-native document format that balances human and AI readability, comparing markdown and HTML, and mentioning LlamaParse for handling existing formats.
Discusses the challenge of incentivizing employees to use internal AI agents instead of traditional support channels, given existing relationships with human helpers.