Runtime-Agnostic AI Workflows: a Pattern for Production Durability and Fast Eval Iteration
Brex’s design keeps agent orchestration identical in production and evals by making it plain business logic over a typed Steps interface, then swapping in a Temporal adapter for durable execution or an in-process adapter for fast tests. That removes eval–production drift rather than relying on discipline.
The trade-off is explicit: runtime-native features must be re-exposed through the portable interface, but the payoff is concrete. Their 20–60 minute agents improved from roughly 96% completion to 99.9% after Temporal replay handled worker failures and redeploys.