Agent Workflow Notes from Deep Research

Deep Research is not a “ask once, answer once” chat box. It is closer to a research workspace: after the user enters a research question, the backend uses LangGraph and LangChain to coordinate search, source filtering, analysis, and report generation. The frontend then presents conversation progress, source lists, reasoning state, and the final report.

The real stack today is Next.js 16, React 19, TypeScript, Tailwind CSS, Zustand, and Supabase JS on the frontend. The backend uses Python 3.11+, FastAPI, Uvicorn, LangGraph, LangChain, OpenAI-compatible APIs, and the Supabase Python SDK.

Why It Needs a Workflow

Open-ended research should not jump straight to a final paragraph. The user needs the process:

• which sources were searched;
• which materials were filtered out;
• why some evidence entered the report;
• whether the final answer can be traced back to sources.

The value of an agent is not just tool calling. The value is making a multi-step process observable and reusable.

Current Design Focus

The README already defines the core goal clearly: Deep Research should turn one complex research task into a traceable, reusable, and presentable workflow. That means the product cannot be designed around only an input box. It has to be designed around the research process.

The things I care about most right now:

1. Research task state should be visible, not hidden behind a generic loading state.
2. Source lists should be part of the report, not a loose appendix.
3. Intermediate reasoning and summaries should stay short enough to review.
4. Observability tools such as Langfuse trace should help debug workflow quality, not only token usage and latency.

For AI agents, the hard part is not generating more text. The hard part is making the generation process recoverable, explainable, and easy to iterate.