agentsop-code-execution-decision

To stop silent arithmetic errors, apply a deterministic gate that routes computable steps to code execution. This prevents LM agents from reasoning through multi-step arithmetic in prose, ensuring verifiable outcomes via Program-of-Thought workflows.

You should use code execution for deterministic transforms like sorting and parsing, and prose reasoning for judgment tasks like tone or synthesis. A gate-based decision policy evaluates each step to route it to the correct mode.

The best way is mixed-step decomposition, which splits tasks into a computable component for code execution and a judgment component for prose reasoning. This ensures deterministic cores run as code while subjective parts use natural language.

Agents waste sandbox runs on judgment tasks due to over-coding, lacking a clear routing rubric. Applying a deterministic gate prevents this by reserving code execution for verifiable computation and using prose for open-ended synthesis.

Common anti-patterns include failing to feed computed results back into the LM, creating unbounded retry loops, and assuming unsandboxed REPL tools are safe. Operational safety guidance enforces result-back requirements and sandbox safety.

Yes, a sandbox is required for safe code execution. The decision policy enforces sandbox safety and result-back requirements, ensuring that deterministic computations run securely without relying on unsandboxed REPL tools.