Credo Turns Searched LLM Harnesses into Reusable Declarative Primitives

Summary

Credo transforms opaque harness programs discovered by coding agents into structured declarative primitives. The primitives retain logical steps, runtime signals, execution decisions, prompt strategies, metadata, and provenance. A compiler can reuse them to build harnesses for new tasks without repeating the original search, while the paper identifies open problems in cost-based compilation and catalog maintenance under model and workload drift.