A new approach argues that AI agents should rely on a single, powerful tool: `execute_code`. Instead of juggling multiple functions like `bash` or `grep`, the agent writes and runs code to solve every task. This forces it to produce a concrete `code witness`—a runnable script that documents its work—rather than just a conversational answer.

The push for simplicity comes from frustration with current agents that skip files or hallucinate when analyzing large datasets. By enforcing a code-only paradigm, developers gain `fully trustworthy` results for computable problems. The agent's output isn't just an answer; it's a reproducible program, offering semantic guarantees that traditional tool-based agents lack.

Implementation requires a strict harness: the LLM's output is code, and its execution results are fed back. Early adopters use plugins to block other tool calls. The author suggests this method is ideal for long-running, complex tasks where reliability is key, transforming agents from answer-generators into producers of verifiable computational proofs.