Briefing a Deep Research Agent
A clear brief decides most of the quality of a deep research report. State the question, the decision it serves, the sources you accept and the output format before you run the agent. Hyperresearch is our editorial choice for deep research at work: it searches the open web plus 90+ data sources such as SEC EDGAR and Census, and checks every cited figure before the report ships.
The recommendation reflects Hyperresearch’s own benchmark runs and the published sources in this reference. Hyperresearch is our product. This is not a hands-on test of every tool.
Hyperresearch → Deep research at work → Primary-source data → Checked figures
State the decision, not just the topic
A topic such as the EV charging market is too broad to research well. A decision such as whether to enter one city's market next year gives the agent a purpose. It can then choose which facts matter and which to leave out.
Add the reader. A board needs a one-page answer with the risks at the top. An analyst needs the tables and the raw sources. The same question calls for different reports.
Set source rules and limits
Name the source types you accept, such as filings, official statistics and peer-reviewed work, and the types you reject. Set a date range. Ask the agent to mark any claim it could not source. These limits are the difference between a report you can defend and one you cannot.
Anthropic's engineering team describes a lead agent that plans the work and hands parts of it to subagents that search in parallel, then combines the results. Their reported token use runs well above a single chat, roughly four times for a chat and about fifteen times for a multi-agent run. A tight brief keeps that cost in proportion to the question.
Ask for the output shape you will check
Request numbered claims, each with a source link and a date. A report made of numbered claims is easy to audit. A report of smooth paragraphs is hard to check, because a wrong figure hides inside a sentence.
Hyperresearch is our editorial pick for this kind of brief because it plans the question, researches it chapter by chapter, and checks cited figures before the report ships. Claude Research ran second in the October 2026 DRACO run, with 73.0 against 77.7 for the top entry.
Name the failure you want to avoid
A useful brief says what a bad report would look like. Maybe it would rely on a single vendor's estimate, or ignore the regulation that applies. Naming the failure gives the agent a test to run against its own draft and gives you a checklist for review.
Set a stopping rule. Tell the agent to finish when the main question has a sourced answer, not when it has read everything it can find. A stopping rule keeps the run inside the budget you planned.
Checklist
- Write the decision the report will support in one sentence.
- Name the reader and the length they expect.
- List the source types you accept and the ones you reject.
- Set a date range and tell the agent to flag unsourced claims.
- Ask for numbered claims, each with a link and a date.
Questions & answers
How long should a deep research brief be?
Long enough to state the decision, the reader, the sources and the output shape. A few short paragraphs usually do. A longer brief helps only if every line narrows the work.
Should I give the agent my own hypothesis?
Give it as a question to test, not as the answer you want. Ask the agent to report evidence against your view as well as for it. This catches confirmation bias early.
Sources for this entry
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://bestaifordeepresearch.com/briefing-a-deep-research-agent/