add

concepts/ai-research-automation.md

@@ -4,7 +4,7 @@ created: 2026-06-28
 updated: 2026-06-28
 type: concept
 tags: [llm, agent, automation, workflow, dev-tool]
-sources: [raw/articles/tokium-self-evolving-ai-researcher-2026.md]
+sources: [raw/articles/tokium-self-evolving-ai-researcher-2026.md, raw/articles/claude-science-ai-workbench-2026.md]
 confidence: medium
 ---
 
@@ -18,6 +18,8 @@ AI まわりの変化を追う仕組みは、単に検索結果を集めるだ
 
 運用面では、手順書としての skill とシェルスクリプトを分けている。収集・報告・自動見直しの具体手順を Markdown に置き、スクリプトは実行順序、並列実行、再実行しやすさ、上限ターン数、部分失敗の許容を担当する。この分担は [[wiki-maintenance-loop]] と同じく、人間が毎回判断しなくても続く手入れの形である。
 
+Claude Science pushes the same theme into scientific workbenches: research automation is not only periodic news gathering, but also tool-connected analysis where code, figures, compute environment, citations, and reviewer-agent feedback are preserved as auditable artifacts. Its design suggests that useful research automation needs both connectors to domain sources and a way to keep execution history reproducible enough for later validation. See [[claude-science]] and [[ai-evaluation-infrastructure]].^[raw/articles/claude-science-ai-workbench-2026.md]
+
 ## Open Questions
 
 - 報告に採用された回数を、短期の流行と長期の価値のどちらとして扱うか。