Tags
4 posts
Three official ways to connect AGENTS.md to CLAUDE.md all loaded the document in testing, with at most 122 tokens of difference. The actual risk is not the wiring but a single example line wrapped in a code block that quietly erases the whole document.
A committed rules file like CLAUDE.md or AGENTS.md is a suggestion the assistant reads, not a rule anyone enforces. When one user request conflicts with a single rule, the model judged the whole file a prompt attack and discarded all of it, 6 times out of 6.
Ten of 33 robots.txt cells let a blocked URL through and Codex clipped the back of a 34 KiB AGENTS.md. All 219 runs exited 0. Which parser gave your verdict?
The first empirical study evaluating AGENTS.md effectiveness has been published. We analyze its impact on coding agent success rates and inference costs.