AI workflows face a fundamental tension: production durability requires heavyweight, persistent runtimes (like Temporal), while LLM eval iteration demands lightweight, ephemeral, in-process loops. Brex's engineering team solved this by decoupling orchestration logic from the runtime entirely. The orchestration is written as a pure function against a typed Steps interface, with no runtime-specific imports. Two adapters then plug in: a Temporal adapter for production (mapping each Steps method to a durable activity) and a minimal in-process eval adapter that swaps real services for fixtures while keeping LLM calls live. This guarantees the exact same orchestration code runs in both contexts, eliminating eval-prod skew by construction. The approach raised long-running agent completion rates from ~96% to 99.9%, though it sacrifices direct access to runtime-native primitives and requires all new capabilities to be threaded through the agnostic interface.
