Four-Gate Test Helps Developers Decide When an LLM Is Actually Needed in Agentic Systems

Agentic systems often place large language model (LLM) calls at every decision node simply because inputs are text-based, resulting in pipelines that are slow, costly, and difficult to test. A four-gate framework — testing enumerability, correct-answer derivability, inter-human agreement, and need for novel synthesis — helps determine whether a node genuinely requires an LLM or can be handled by deterministic code. Passing even one of the four gates means a rule-based or algorithmic solution is preferable, and in practice most nodes fail at gate one or three. Beyond the four gates, a second axis examines whether a node's output can be verified deterministically after the fact, which affects how much an unreliable model judge can be tolerated. The approach, practiced at infrastructure firm Pulsed Media, advocates keeping the LLM surface small — a few true judgment nodes built on a large deterministic base — rather than scattering model calls throughout a pipeline.