Checking a compiled workflow for faithfulness does most of the reliability work

The authors' starting claim is that a natural-language workflow handed to an agent is not yet software: it carries no explicit data-dependency or control-transfer declarations, so an agent can drift from what was actually specified without anyone noticing until the output is wrong. Their proposed fix compiles the workflow into an artifact-driven form, refines it with constrained optimization, and checks it for faithfulness by decomposing it into local obligations and running scenario-based dry runs before execution. Tested across 488 instances spanning 11 domains, an ablation that isolated the faithfulness-validation step on its own found it responsible for a 16 percentage point improvement.

Why it matters: For anyone specifying agent workflows in plain language and hoping the agent does what was meant, this is a concrete alternative to hoping: compile the spec, then check the compiled form for faithfulness before running it. The ablation result matters because it says where the gain actually comes from. Most of it is the validation step, not the act of compiling on its own.