AI evidence extraction

From Specification to Execution: AI Assisted Scientific Workflow Management

Komal Thareja, Hamza Safri, Rajiv Mayani, Anirban Mandal, Ewa Deelman · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Case study with comparative evaluation.

Primary method

Design science with specification-driven development; system architecture design integrating multiple components (AI authoring, debugging agent, Pegasus WMS, MCP interface)

Main result

The system generated and executed large-scale workflows with thousands of jobs, reduced debugging effort, and allowed non-expert users to construct workflows with expert-level design patterns. "The skill-based approach brings non-expert users closer to expert-level workflow design, reducing the need for manual expertise but leaving room for expert refinement." Claude produced a production-ready workflow in 2 sessions (52 prompts, ~$10–15) with superior reproducibility compared to Codex and Kimi, which required multiple sessions and exhibited convergence issues with Pegasus-specific staging conflicts.

Research paradigm

Design science / pragmatist

Author conclusions

"We presented an AI-assisted approach to scientific workflow management that combines specification-driven workflow generation, automated debugging, hierarchical execution with Pegasus WMS, and MCP-based remote management." The authors conclude that "the system could generate and autonomously execute workflows with thousands of jobs" and that "workflow structure, in particular the number of rounds, is the dominant factor in execution cost." They propose future work on "evaluating the performance of the MCP layer, and building an AI-driven platform that guides users in designing more complex workflows, and manages the full workflow lifecycle."

Risk of bias

Selection bias in use case choice (federated learning may not represent all workflow types); Comparison bias: manual baseline developed by single expert may not represent typical manual workflow development; Funding bias: authors acknowledge NSF support but work is on their own system; LLM selection bias: Claude, Codex, and Kimi represent different model capabilities and cost structures; Comparison bias: Claude Code is tested with structured pegasus-ai plugin skills, while Codex and Kimi only have static documentation context—unequal conditions for comparison; generalizability to other scientific workflow types unclear; LLM version differences: Different LLM versions used (Claude Opus 4, GPT-5.4, Kimi K2.6) with different training dates and capabilities; Small-scale data: TCIA and NIH chest X-ray datasets may not reflect production federated learning scenarios; Evaluation limited to specific AI models (Claude, Codex, Kimi); results may not generalize to other LLMs; may reflect individual preferences rather than general best practices; No blinded evaluation of generated vs. manual workflows

Limitations

The authors explicitly state: "A quantitative evaluation of the MCP layer, including gateway routing behavior, multi-node coordination, and remote-client interaction patterns, is outside the scope of this paper." Additionally, the federated learning performance gap is attributed to data constraints: "the effective data available per client in the federated setting is limited
For example, in the TCIA setup with K=10 clients, each client receives approximately 330 samples, which constrains local training capacity." The manual workflow baseline represents a single expert implementation, and the study does not systematically evaluate the MCP-based remote management layer quantitatively.

Open questions raised

Future work will focus on evaluating the performance of the MCP layer, and building an AI-driven platform that guides users in designing more complex workflows, and manages the full workflow lifecycle - from specification to optimization of workflow runs across heterogeneous infrastructures.
Quantitative evaluation of the MCP (Model Context Protocol) layer, including gateway routing behavior, multi-node coordination, and remote-client interaction patterns
Extension to workflow types beyond federated learning (generalizability of the approach)
Exploration of advanced federated learning design choices such as early stopping and alternative orchestration strategies (e.g., Ensemble Manager)
Integration of federated learning-specific optimizations into the AI-assisted generation system

Data: TCIA (The Cancer Imaging Archive): 4,144 3D CT/MRI lung volumes, publicly available; NIH ChestX-ray: 50,000 2D chest X-ray images, publicly availableCode: https://github.com/pegasus-isi/claude-plugin-marketplace/tree/main/plugins/pegasus-ai; https://github.com/pegasus-isi/medical-imaging-fl-workflow (Claude-generated workflow)Extracted from: pdf

Explore related topics

AI-Assisted Discovery Research Productivity Design Science