The latest articles on DEV Community by Gongsheng Li (@alcholiclg). https://dev.to/alcholiclg https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3634824%2Fb78ccc79-bb30-4ba1-83ca-90a190e43421.jpeg https://dev.to/alcholiclg en Gongsheng Li Fri, 28 Nov 2025 14:41:19 +0000 https://dev.to/alcholiclg/rethinking-my-deep-research-agent-workflow-should-we-move-beyond-static-trees-2eac https://dev.to/alcholiclg/rethinking-my-deep-research-agent-workflow-should-we-move-beyond-static-trees-2eac <p>I’m reevaluating a deep-research workflow I built earlier and would love some advice.</p> <p>My previous design used a static tree workflow (fixed width/depth, node = search → extract → summarize → generate follow-ups), similar to GitHub’s popular deep-research repo. But newer projects like deer-flow and open_deep_research seem to favor a different style:<br> clear multi-agent roles + dynamic tool-call loops instead of a fixed search tree.</p> <p>I’m trying to understand:</p> <ol> <li>Is moving from static workflows to tool-call loops the current trend? What are the concrete advantages, and is it worth refactoring?</li> <li>How do you evaluate these systems? From output alone it’s hard to tell which is “better,” and static workflows are still very popular. Is there actually a meaningful performance gap today?</li> <li>For a practical open-source project, what principles guide iteration? If the goal isn’t just scoring well on benchmarks (e.g., HLE), how would you think about evolving a deep-research agent? Any thoughts or experience would be really helpful. Thanks!</li> </ol> ai rag programming