Token "Optimizers" for AI Coding Agents Are Silently Dangerous (old.reddit.com)

The guardrail point resonates. In practice, I’ve found context selection matters as much as model choice—especially when the agent needs to avoid editing outside the requested scope.