The business
Whitfield Research Group, based in New York, sells sector reports to investors and strategy teams. Analysts used to spend about three days per report. A multi-agent system now drafts reports in hours, and analysts review them.
The system
A coordinator agent receives the research question and delegates to four subagents built with the Agent SDK:
| Subagent | Job | Tools |
|---|---|---|
| Web search | Finds and reads public sources | web_search, fetch_page |
| Document analysis | Reads client-supplied PDFs and datasets | read_document, extract_tables |
| Synthesis | Combines findings into a structured draft | none of its own today |
| Report | Formats the final report with citations | render_report |
Every subagent currently receives the full catalogue of 16 tools.
What goes wrong today
- Simple questions ("When did the EU battery regulation take effect?") still run all four subagents. A simple answer costs about the same as a full report.
- Reports on broad topics cover some subtopics in depth and skip others.
- Citations disappear between the search results and the final report.
- Two sources sometimes give different figures for the same market; the draft quietly picks one.
- When the web search subagent times out, the coordinator receives "search unavailable" and the report says nothing about the gap.
- Synthesis input runs to about 60,000 tokens of raw excerpts, and the middle subagents' findings are under-represented in drafts.
Standards for a finished report
- Every factual claim traces to a source with a URL or document name, and a date.
- Conflicting figures are shown side by side with their sources.
- A coverage section lists what could not be researched and why.
People
| Stakeholder | Cares most about |
|---|---|
| Head of Research (sponsor) | Reports analysts can trust and sign |
| Senior analyst | Coverage and correct citations |
| ML engineer | Cost per report and debuggable runs |