Use Deterministic Code Before Calling an LLM in AI Agents

A common assumption in AI application design is that an LLM should handle every decision, but this can be wasteful. The article argues that recurring problems with known solutions can be converted into ordinary deterministic functions. In an example of an agent analyzing failed CI/CD builds, a repeated NuGet error can be resolved by simple pattern-matching code instead of a fresh model call. The author recommends a division of labor where code handles known patterns and explicit rules while the LLM tackles ambiguous or unfamiliar cases. Deterministic logic also offers the benefits of lower inference costs, faster responses, and predictable behavior.