Mine your agent's repeated traces into deterministic workflows — and let the compiler refuse when side effects are irreversible

arXiv 2608.02680·medium signalXBlueskyLinkedInCopy link

TraceCompiler clusters noisy agent traces for the same intent and compiles them into mostly-deterministic workflows, classifying every data binding into constants, user inputs, copied outputs, transforms, or residual LLM decisions so you can see exactly how much model judgment survives. Producer-consumer dependency recovery hits 0.928 precision / 0.943 recall on 15,775 def-use edges, far above adjacency (0.711 F1) and frequency-thresholded directly-follows (0.712 F1), with 0.993 precision on AppWorld. The design detail worth stealing is the refusal: a Venmo money-request intent compiled from 34 API calls down to 11 runtime calls, while a Spotify/Todoist intent was correctly refused because its irreversible side effects were under-determined — and the authors explicitly decline to claim net efficiency since offline compilation cost went unmeasured.