Artic Compiles Natural-Language Workflows Into Artifact-Declaring Steps and Lifts Agent Task Resolve Rate 28 Points Across 11 Real Domains
The authors argue natural-language workflows fail as agent instructions because data dependencies stay implicit, forcing the executor to guess which prior results a step should read, and because agents drop long or branching instructions under context pressure. Artic compiles the workflow so each step declares the artifacts it reads and writes, constraints gate produced artifacts, and control transfers are explicit, which exposes steps that depend on too much state and lets the compiler refine them through constrained optimization. Faithfulness of the LLM-assisted transformation is checked by decomposing it into local obligations plus scenario-based dry runs, and across 488 problem instances from 11 real-world domain workflows the compiled version improved task resolve rate by 28 percentage points over the original text.
↳ Follow the thread