AI agents can now execute code, call APIs, and run tools.

But most agent architectures still look like:

Agent Tool Execution

Which raises a question: