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

A research team has introduced the Source Code Agent framework, a new architecture that decouples large language model workflow logic from the generative model itself to achieve deterministic, verifiable execution in structured operational environments [1]. The framework is built on a "Blueprint First, Model Second" philosophy. An expert-defined operational procedure is first codified into a source code-based Execution Blueprint, which is then run by a deterministic engine. The large language model is invoked only as a specialized tool to handle bounded, complex sub-tasks and is never permitted to decide the workflow's path [1][2]. Yuhang Ye and collaborators evaluated the approach on the TravelPlanner benchmark for constraint-aware travel planning. The Source Code Agent achieved a 35.56% final pass rate, a 97.6% improvement over the state-of-the-art ATLAS baseline, which recorded an 18.00% pass rate on the same Claude-Sonnet-4 backbone [1][2]. Constraint violations fell by 96.0%, dropping from 275 in the ATLAS baseline to 11 in the Source Code Agent framework [1][2]. Execution efficiency improved by 27.1%, with the framework requiring an average of 10.2 steps with a standard deviation of 0.7, compared to 14.0 steps for ATLAS [1][2]. The authors argue that current LLM agent architectures conflate probabilistic, high-level planning with low-level action execution inside a single generative process, which limits their use in settings where procedural fidelity and predictable execution are strict requirements [1][2]. By separating the workflow blueprint from the model's generative capabilities, the framework targets applications governed by strict procedural logic. Two production incident-diagnosis deployments and additional results on the ScienceWorld and ALFWorld benchmarks indicate the architecture transfers beyond travel planning to other procedurally well-defined, constraint-intensive workflows [1][2].