Workshops Archive

Agents at Your Service: Reuse-First Workflow Composition with Laminar 3.0


Workshop: 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26)

Authors: Marco Edoardo Santimaria (Università di Torino, Italy); Rosa Filgueira (Edinburgh Parallel Computing Centre); and Alberto Mulone, Iacopo Colonnelli, and Marco Aldinucci (Università di Torino, Italy)

Abstract: Large language models show promise as software development assistants, but their tendency to hallucinate makes them unreliable for directly generating workflow implementations in specialized coordination languages. This paper presents Laminar 3.0, an evolution of the Laminar serverless framework that integrates a bounded, prompt-chained pipeline of role-specialized LLM agents to support workflow discovery and composition. It introduces an LLM-driven metadata generation mechanism to enrich the workflow registry, a hybrid retrieval mechanism combining semantic search, lexical search, and LLM-based re-ranking, and a retrieval-first, constraint-guided generation pipeline that proposes workflow skeletons only when no suitable artifacts exist. Laminar 3.0 prioritizes reuse and synthesizes only the coordination layer, leaving host semantics under developer control. Experiments show semantic retrieval remains robust even for compressed queries, achieving 79% top-1 and 81% top-3 retrieval, while the generation pipeline produces skeletons with a CodeBLEU score of 0.946, substantially outperforming direct generation with vanilla prompting.


Back to 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26) Archive Listing

Back to Full Workshop Archive Listing