REFLEX assigns bounded choices, such as tool selection or continuation, to Jev, a fast typed decision model. It executes high-confidence choices directly and escalates low-confidence decisions or free-form generation to a strong LLM.
On a frozen 100-task benchmark, REFLEX reached 95% success while making 72.7% fewer strong-model calls than a strong-only agent. Reductions persisted across three fallback families. Reliability declined with larger action sets and near-valid alternatives around authorization boundaries, while external evaluations found little advantage over a cheap generative cascade when ordinary routing was already accurate.