MOSAIC: Modular Orchestration for Structured Agentic Intelligence and Composition

Researchers introduce MOSAIC, a structured agentic framework that automates data science model selection by combining LLM flexibility with systematic verification. Unlike traditional AutoML systems or unstructured LLM agents, MOSAIC creates intermediate 'blueprints' that ground decisions in retrieved evidence and execution feedback, improving task performance and decision traceability.

MOSAIC addresses a fundamental limitation in current AI automation: the gap between rigid AutoML pipelines and flexible but opaque LLM-based agents. The framework treats automated data science as a staged, structured search problem rather than unconstrained synthesis, introducing blueprints—intermediate representations that specify modeling components, composition rules, and execution requirements. This approach transforms how LLMs generate code by grounding suggestions in retrieved prior cases and source-code modules rather than relying on token prediction alone.

The research builds on decade-long progress in AutoML systems like Auto-sklearn and recent advances in agentic AI. However, MOSAIC's innovation lies in hybrid architecture: it leverages LLM flexibility while maintaining verifiability and reusability through systematic composition. The framework validates candidates through execution and refines them using diagnostic feedback and reinforcement learning—critical for financial applications where failures carry real costs.

The financial time-series use case demonstrates practical relevance. Models must satisfy multiple criteria simultaneously: predictive accuracy, distributional fidelity, execution reliability, and downstream financial metrics like risk and tail behavior. This mirrors real-world constraints where academic performance metrics diverge from operational requirements. MOSAIC's staged refinement using task-specific feedback creates a feedback loop absent in standard approaches.

For AI development, MOSAIC signals growing maturity in agentic systems. Rather than pursuing unconstrained autonomy, the research prioritizes structured decision-making and traceability—requirements for enterprise and regulated domains. This validates a design philosophy where agency works within intelligible constraints, potentially influencing how production AI systems integrate LLMs with domain expertise and execution feedback.

• →MOSAIC introduces intermediate 'blueprints' that structure LLM-based model selection, improving traceability and reusability compared to unstructured agentic approaches.
• →The framework grounds code generation in retrieved evidence from prior cases rather than unconstrained synthesis, reducing hallucination and improving execution success.
• →Validation through execution and refinement using diagnostic feedback enables MOSAIC to handle multi-criteria optimization essential for financial applications.
• →Hybrid design combines AutoML rigor with agentic flexibility, suggesting evolution toward verifiable AI systems rather than pure black-box approaches.
• →Financial time-series experiments demonstrate applicability to high-stakes domains where execution reliability and downstream metrics matter beyond accuracy scores.