CURATE framework uses LLM agents to automate reproducible workflow lifecycle

CURATE (Composition, User-in-the-loop, Reuse, and Automated Task Execution) is a new multi-agent framework designed to address a critical gap in agentic code generation: existing coding agents focus on generating and testing code but ignore the rest of the workflow lifecycle—deployment, sharing, and reuse. Researchers from the arXiv paper (submitted Aug 2026) propose a human-in-the-loop system where LLM agents handle everything from composing modular workflows to executing tasks, with a key catalog component that stores modules for reuse across different projects. This catalog builds a foundation for FAIR principles (Findable, Accessible, Interoperable, Reusable) by allowing curated modules and subgraphs to be shared, which is especially valuable for scientific computing.

To prove the concept, the team built an initial prototype using Anthropic's Claude Opus 4.8 and ran 6 experiments. They reproduced and adapted 4 workflows from the SeBS-Flow benchmark suite (a standard for serverless scientific workflows) and automated the development and scaling of an environmental engineering workflow that simulates anaerobic digestion—a complex mechanistic model. The results demonstrate that CURATE can not only generate code but also deploy it, manage dependencies, and store components in a catalog for later reuse. This effectively moves LLM agents from code generators to workflow orchestrators, potentially lowering barriers for scientists and engineers who need reproducible, shareable pipelines.