MOSAIC: Multi-agent Orchestration for Task-Intelligent Scientific Coding
Abstract
We present MOSAIC, a large language model (LLM) based multi-agent framework for tackling complex scientific coding tasks. Unlike general-purpose programming, scientific coding requires rigorous algorithmic reasoning, substantial domain expertise, high numerical precision, extended context management, and the ability to decompose and solve interdependent subproblems under a larger objective. To address these challenges, we design and integrate specialized agents responsible for self-reflection, planning, coding, and debugging inspired by a teacher–student paradigm. This architecture balances open-ended exploration, ensuring executability while maintaining a consolidated context to minimize hallucinations. We evaluate MOSAIC on scientific coding benchmark and show that our framework, together with its specialized agents, outperforms existing approaches in accuracy, robustness, and interpretability.