AGENTS.md updated with generic, software-oriented agent rules to address LLM behavior

This burns about ~3K context, in exchange it seems to address some of the bad behavior from LLMs; your mileage may vary. It's a combination of my earlier output style and an updated generic rule set for agents to...