Letting the model write a program and run it. It removes a whole class of weaknesses at once, since arithmetic, data manipulation and file processing are things code does reliably and language models do not. It also means untrusted code is running somewhere, which needs taking seriously.
Technical view
Executed in a sandboxed interpreter with restricted network and filesystem access, resource limits and a timeout, returning stdout, stderr and produced artefacts to the model for inspection and retry. The generate-run-observe-fix loop substantially improves reliability on quantitative tasks. Security requirements are the same as for running any untrusted code, and the additional risk is that the code is authored by a system that reads untrusted input.
Related
Part of AI Terminology Index, section: Agents and Tools.