> I don't like it but don't know what to do, Anthropic models especially increasingly ignore instructions whether in memory or agents files.

I've taken to instructing the agent to manage the subagent, and the principal agent's sole job is to ensuring the subagent follows instructions to the letter.