#Code-Execution

The Code-Only Agent

The article argues that an agent should have only one tool, code execution. By forcing every task into executable code, it produces repeatable witnesses, clearer semantics, and more trustworthy computation.