Running model-generated commands or programs under controlled authority.
Code execution gives an agent a powerful general-purpose tool for calculation, transformation, testing, and automation. It must be isolated, resource-limited, observable, and separated from sensitive credentials and production authority.