AI-assisted system generates and executes scientific workflows via MCP

Komal Thareja, Hamza Safri, and Rajiv Mayani present an end-to-end AI-assisted framework for scientific workflow management that addresses a core limitation of existing LLM-based approaches: direct code synthesis, which the paper argues limits transparency, reproducibility, and integration with established workflow systems. Their method introduces a structured specification phase that explicitly separates workflow intent, design, and implementation, enabling validation before code generation begins. This separation is intended to make the system's reasoning auditable and to reduce errors that arise when LLMs jump straight to implementation.

The framework also includes an LLM-based debugging agent capable of diagnosing and resolving failures across multiple system layers, reducing the manual effort typically required to troubleshoot complex pipelines.

The framework also includes an LLM-based debugging agent capable of diagnosing and resolving failures across multiple system layers, reducing the manual effort typically required to troubleshoot complex pipelines. For distributed execution and user interaction, the team integrates Pegasus — a widely used scientific workflow management system — with a Model Context Protocol (MCP) layer, creating a unified interface for workflow submission, monitoring, and control.

The approach was evaluated using a federated learning workflow for medical imaging, selected for its parallel, iterative, and dependency-intensive structure. The system successfully generated and executed large-scale workflows containing thousands of jobs, reduced debugging effort, and allowed non-expert users to construct workflows with expert-level design patterns. The authors conclude that end-to-end AI-assisted workflow generation and execution is feasible and point toward AI-driven platforms for managing the full scientific workflow lifecycle.

Key facts

1. 01Authors: Komal Thareja, Hamza Safri, Rajiv Mayani
2. 02The system introduces a structured specification phase that separates workflow intent, design, and implementation before code generation
3. 03An LLM-based debugging agent diagnoses and resolves failures across multiple system layers
4. 04Pegasus WMS is integrated with a Model Context Protocol (MCP) layer for workflow submission, monitoring, and control
5. 05Evaluated on a federated learning workflow for medical imaging with thousands of jobs
6. 06Non-expert users were able to construct workflows with expert-level design patterns
7. 07The paper argues direct code synthesis limits transparency, reproducibility, and WMS integration

Topics