I think part of the problem is that current AI tooling massively lowers the cost of making infrastructure and product changes, but doesn’t equally improve consequence reasoning.
Generating changes is becoming cheap. Predicting second and third order effects across large systems is still hard.
a native support for AGENTS.md would be huge as it's already the standard every AI coding tool reads for project context but right now it lives in the repo but github does nothing with it, but if github surfaced it can be machine readable project policies