Developer finds AI coding agents routinely ignore their own stored rules

A developer ran controlled tests across hundreds of sessions to measure whether a coding agent actually follows rules loaded into its memory, finding the results far worse than expected. Despite relevant rules being retrieved and displayed to the model, the agent repeated the same heredoc-related coding mistake four times in a single conversation. Tests showed the agent followed a pointer instruction to read a rules file in only 4.9% of 89 sessions, and even with all 59 rules loaded simultaneously, violations still occurred six times in one day. The author found that neither preloading rules nor fetching them on demand could reliably stop a model once it chose to ignore a rule, and that enforcement required hook-based checks at the tool-call or turn-completion stage. The investigation also revealed that rule compliance varied significantly across different agent hosts and that many rules lack observable outputs, making automated enforcement structurally difficult.