Deep research agent that writes cited reports from web and local docs

What they did

They stood up a deep research workflow on top of GPT Researcher, the open deep research agent that handles both web and local research for any given task. The agent takes a research brief, decomposes it into sub-questions, runs parallel searches across web sources and local documents, and then synthesizes the findings into a detailed, factual, unbiased report with citations. They used AG2 as the orchestration layer to coordinate the planner, the retrieval steps, and the writer, so the pipeline could be extended with additional agents or custom steps. The output was a long-form report with inline citations rather than a chat answer, which made it usable as a first draft for internal briefs, competitive scans, or literature-style summaries. They also leaned on the project's customization options to tune the report structure and source handling, and ran it both from the Python package and via the provided Docker/Colab paths to compare local versus hosted execution.

How it went

The pipeline produced structured, citation-backed reports that read like a solid first draft, cutting the blank-page phase of research work. Parallel sub-question search kept latency reasonable for multi-source briefs, and the AG2 layer made it straightforward to add steps. Polish level is prototype-to-internal-tool: reports still need a human pass for source quality and claim accuracy before anything external.

Worth knowing

Report quality depends heavily on the underlying model and search provider, and citations can point to weak or tangential sources that need manual vetting. Long briefs cost real tokens and time, so budget and rate limits matter. Local document research requires wiring your own document store and embeddings, and the README excerpt does not cover that setup in detail. No built-in fact-checking pass, so treat outputs as drafts.

Wiring

Install the gpt-researcher package (or clone the repo) and set API keys for the chosen LLM and search provider in a .env file. AG2 sits on top as the orchestration layer: define a planner agent that emits sub-questions, a retrieval step that calls GPT Researcher's search/scrape pipeline for each sub-question (web and local doc retrievers configured separately), and a writer agent that receives the aggregated findings and drafts the cited report. Configure report type, report length, and source count through GPT Researcher's config, then pass the brief through the AG2 group chat or sequential pipeline. Output is written to a markdown file; optionally run via the provided Docker image or Colab notebook for a hosted path.

Try it yourself

Set up a deep research agent in this repo using GPT Researcher and AG2. It should accept a research brief, break it into sub-questions, search both the web and any local docs I point it at, and produce a long-form markdown report with inline citations and a source list. Add a config file for model choice, search provider, and max sub-questions, plus a CLI entry point that takes a brief and writes the report to ./reports. Include a short README section on how to add local documents.