Safely Executing LLM-Proposed Actions with Typed Verifiers

This article presents a pattern for safely executing actions proposed by large language models (LLMs) by using a deterministic verifier and a strictly typed executor, avoiding direct execution of LLM output.

Details

The article discusses the problem of directly executing the probabilistic output of LLMs, which can lead to accidents. It proposes a pattern with three key components: 1) Structuring the input schema to define what counts as grounds, 2) Having the LLM output a plan of typed actions instead of free text, and 3) Using a deterministic verifier to accept, reject, or degrade the plan before the executor runs the verified typed actions. This approach aims to make the system more operable and reduce the risk of accidents from executing LLM output directly. The article also explores a thought experiment around just-in-time production access, where the LLM proposes a plan but a compiler-like verifier determines eligibility, least privilege, and approval before the executor runs the approved actions.