Rethinking My Deep-Research Agent Workflow — Should We Move Beyond Static Trees?

I’m reevaluating a deep-research workflow I built earlier and would love some advice.

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:
clear multi-agent roles + dynamic tool-call loops instead of a fixed search tree.

I’m trying to understand:

1. Is moving from static workflows to tool-call loops the current trend? What are the concrete advantages, and is it worth refactoring?
2. 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?
3. 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!