Blueprint First, Model Second: A Framework for Deterministic LLM Workflow

The article introduces the \textsc{Source Code Agent} framework, which employs a "Blueprint First, Model Second" approach to enhance the deterministic execution of large language models (LLMs) in structured environments. By separating workflow logic from generative processes, this framework allows for expert-defined Execution Blueprints to guide operations, resulting in a 35.56% pass rate on the TravelPlanner benchmark, a significant improvement over the ATLAS baseline. This architecture notably reduces constraint violations by 96% and improves execution efficiency by 27.1%, making it applicable for reliable autonomous agent deployment in procedural and constraint-heavy tasks.