SciVisAgentSkills: Design and Evaluation of Agent Skills for Scientific Data Analysis and Visualization

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, long

Summary

SciVisAgentSkills is a collection of reusable agent skills designed to augment general-purpose coding agents for scientific data analysis and visualization. This collection encodes environment assumptions, tool usage patterns, and domain heuristics across scientific tools such as ParaView, napari, VMD, and TTK. The skills were evaluated on Codex and Claude Code using SciVisAgentBench, a benchmark comprising 108 expert-designed multi-step tasks. Results demonstrate that incorporating these agent skills consistently improves mean task scores across the evaluated suites. While performance improved, token-efficiency benefits varied depending on the agent harness and specific tool setting. These findings underscore the critical role of structured procedural knowledge in enabling reliable, long-horizon SciVis workflows and emphasize that agent skills must be studied in conjunction with their execution harness.

Key takeaway

For Machine Learning Engineers developing agentic scientific visualization systems, integrating domain-specific agent skills is crucial for enhancing workflow reliability and efficiency. You should adopt or create skills that encode tool-specific knowledge and best practices for tools like ParaView or napari. Critically, evaluate how these skills interact with your agent's execution harness, as this impacts token usage and overall performance. This approach offers a low-cost entry point to improve complex scientific tasks.

Key insights

Agent skills augment general coding agents with domain-specific procedural knowledge for scientific data analysis and visualization.

Principles

• Structured procedural knowledge significantly improves long-horizon SciVis workflows.
• Skill effectiveness and token efficiency are interdependent with the agent harness.
• Skills are most impactful for specialized or fragmented domain knowledge.

Method

SciVisAgentSkills are built by fixing tool versions, distilling official documentation, reusing SciVis agent exemplars, and incorporating empirical failure fixes.

In practice

• Implement skills for tools like ParaView, napari, VMD, and TTK.
• Encode environment assumptions and API usage patterns.
• Integrate executable code snippets for common workflows.

Topics

• Agent Skills
• Scientific Visualization
• Coding Agents
• SciVisAgentBench
• ParaView
• Long-horizon Workflows

Code references

• KuangshiAi/SciVisAgentSkills

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• SciVisAgentSkills: Design and Evaluation of Agent Skills for Scientific Data Analysis and Visualization — SciVisAgentSkills augment coding agents with structured procedural knowledge for scientific visualization, improving task performance and efficiency.
• From Research Question to Scientific Workflow: Leveraging Agentic AI for Science Automation — An agentic AI architecture automates scientific workflow generation from natural language, ensuring deterministic execution via layered design.
• Soft Forks: How Agent Skills Create Specialized AI Without Training — Agent Skills enable specialized AI behavior at runtime through context injection, offering a cost-effective alternative to fine-tuning.
• Beyond Prompting: Using Agent Skills in Data Science — Skills package instructions and resources for LLMs, enabling reliable automation of recurring data science workflows.
• On Accelerating Grounded Code Development for Research — Lexical search and structured workflows can effectively ground coding agents in dynamic, domain-specific research.
• Agent Skills — AI coding agents require explicit workflow-based "skills" to enforce senior engineering practices and prevent shortcuts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.