Why did your documentation break when agents started using it?

The files hadn't changed. The content was accurate. Your team had been using the same documentation for months without problems.

Then agents started consuming it — and something broke. Errors appeared where there hadn't been errors. Steps got missed. Outputs contradicted documentation the team considered authoritative.

If that feels strange, you're not imagining it. The documentation didn't get worse. The consumer changed.

What's actually happening

Documentation is designed with a specific consumer in mind — usually a human reader who brings context, infers meaning, and exercises judgment about what applies to their situation. Human readers recover from gaps. They recognize when something doesn't apply. They bring organizational knowledge to everything they read.

AI agents consume documentation differently. They consume it literally, without the ambient context human readers supply. What was written as guidance becomes instruction. What was implicit becomes operational. What was context becomes fact.

The assumptions embedded in your documentation — about what the reader already knows, how they'll handle ambiguity, what they'll infer — were always load-bearing. Human readers were quietly absorbing the failures. When agents became the consumer, that tolerance disappeared.

The pattern has a name: Machine-Consumed Infrastructure.

Why fixing the docs may not fix the problem

The instinct is to improve the documentation — add more detail, fix inaccuracies, make things clearer. That's often the right response to documentation problems.

But when the failure is caused by consumer substitution, improving the content doesn't address the structural issue. Better retrieval returns the wrong assumptions more efficiently. More detail gives agents more context to apply incorrectly.

The question isn't whether the content is accurate. It's whether the design assumptions — about what the consumer already knows, how they'll handle ambiguity, what they'll infer — still hold for the consumer actually using it.

The infrastructure isn't broken. It's being used by a consumer it wasn't designed for. The fix requires reconsidering design assumptions, not just content quality — because the consumer's relationship to ambiguity, context, and inference is fundamentally different from the one the documentation was built for.

Read the framework