SciVisAgentSkills boosts coding agents for scientific visualization
Researchers introduce SciVisAgentSkills, a collection of reusable agent skills that augment coding agents like Codex and Claude Code for scientific data analysis and visualization tasks involving tools such as ParaView, napari, VMD, and TTK.
Score breakdown
The work demonstrates that structured procedural knowledge in the form of reusable agent skills can improve coding agent performance on complex, multi-step scientific visualization tasks where general-purpose agents otherwise lack tool-specific expertise.
- 01SciVisAgentSkills is a collection of reusable agent skills for scientific data analysis and visualization.
- 02Skills encode environment assumptions, tool usage patterns, and domain heuristics for ParaView, napari, VMD, and TTK.
- 03Evaluated on Codex and Claude Code using SciVisAgentBench, a benchmark of 108 expert-designed multi-step tasks.
Kuangshi Ai, Haichao Miao, and Kaiyuan Tang introduce SciVisAgentSkills, a collection of reusable agent skills aimed at augmenting coding agents for scientific data analysis and visualization. While general-purpose coding agents demonstrate strong capabilities, the paper argues they lack the tool-specific expertise needed for scientific visualization (SciVis) tasks. SciVisAgentSkills addresses this by encoding environment assumptions, tool usage patterns, and domain heuristics across several scientific tools: ParaView, napari, VMD, and TTK.
The skills are evaluated on two agents — Codex and Claude Code — using SciVisAgentBench, a benchmark comprising 108 expert-designed multi-step tasks.
The skills are evaluated on two agents — Codex and Claude Code — using SciVisAgentBench, a benchmark comprising 108 expert-designed multi-step tasks. Results show that incorporating agent skills improves mean task scores across the evaluated suites, with token-efficiency benefits that depend on the specific agent harness and tool setting. The authors highlight that structured procedural knowledge is important for enabling reliable, long-horizon SciVis workflows, and emphasize that the effectiveness of skills is tied to the execution harness that loads and applies them. The skills are publicly available on GitHub.
Key facts
- 01SciVisAgentSkills is a collection of reusable agent skills for scientific data analysis and visualization.
- 02Skills encode environment assumptions, tool usage patterns, and domain heuristics for ParaView, napari, VMD, and TTK.
- 03Evaluated on Codex and Claude Code using SciVisAgentBench, a benchmark of 108 expert-designed multi-step tasks.
- 04Agent skills improve mean task scores across the evaluated suites.
- 05Token-efficiency benefits vary depending on the agent harness and tool setting.
- 06The authors find that skills must be studied alongside the execution harness that loads and applies them.
- 07The skills are publicly available at https://github.com/KuangshiAi/SciVisAgentSkills.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →