Briefing a Deep Research Agent: How to Get a 20-Source Report Worth Reading

Overview

A deep research agent is a model you commission rather than chat with: you give it a brief, it spends several minutes to an hour running its own searches across many sources, and it hands back a long cited report. This module is about writing the brief — the part that decides whether the report is worth your time.

Why now. In April 2026 Google introduced Deep Research Max, an agent that uses what its announcement calls “extended test-time compute to iteratively reason, search and refine the final report,” built for asynchronous background runs — the announcement’s own example is “a nightly cron job triggering the generation of exhaustive due diligence reports.” Commissioning research overnight only pays off if the brief was right before you went to bed.

What you’ll be able to do. Scope and brief a deep-research run so the report comes back useful on the first pass, and recognise the one thing you must never skip on the way in.

The content

The obvious read is that a deep research agent saves you the work of being precise — it reads hundreds of pages so you don’t have to, so a rough prompt should be fine. Overturn it: a chat model lets you course-correct turn by turn, but a research agent commits to your brief and disappears for half an hour. There is no turn-by-turn. The run is long and asynchronous, so the consequence is blunt: you need the prompt right the first time.

That changes the unit of work. You are not prompting; you are writing a commission — an objective, a scope, and the constraints that would change the shape of the answer. OpenAI’s deep research may ask clarifying questions to confirm intent before it runs, and the single highest-leverage move is to treat those questions as the real brief: answer every one fully, because each unanswered question is a gap the agent fills with a guess. Name your audience, your time frame, what counts as in and out of scope, and the specific terms — brands, technical names, regions — that the model will search on. Vague in, unfocused out: ask for everything and you get a report that covers everything and decides nothing.

There is a balance to hold, and it is the credible part. OpenAI’s “three tips” guidance flags over-instructing as the bigger trap: pile on rigid constraints and you can “limit creativity and reduce the accuracy and depth of insights.” Give the agent a sharp objective and the freedom to find the route to it.

Now the limit. These agents are built to synthesise the open, indexed web — so they are weakest exactly where the good answer isn’t there: paywalled filings, your internal data, fast-moving stories where the web is still wrong. A confident twenty-source report on a thin evidence base is more dangerous than no report, because the citations make it look settled. The fix is iteration, not faith: read the first run as a draft, see which sources it leaned on, then re-commission with the gaps named.

Try it

Use this on a real research task you’d otherwise have skimmed in browser tabs — a market scan, a vendor comparison, a literature pull. Paste your filled-in brief into a deep research agent and answer its clarifying questions before letting it run.

I want to commission a deep research report. Before you run anything, read my
brief below and ask me any clarifying questions needed to scope it — then wait
for my answers.

Objective (the one decision this report should inform):
[e.g. which of N vendors to shortlist for X]

Audience and use: [who reads this and what they'll do with it]
Time frame / recency: [how current the evidence must be]
In scope: [the questions you DO want answered]
Out of scope: [what to ignore, so it doesn't sprawl]
Key terms to search on: [brands, technical terms, regions, product names]
Source bar: [e.g. prefer primary sources; flag anything thinly evidenced]

For each major claim, cite the source and note how confident you are. At the end,
list the gaps where evidence was weak or missing — I'll use those to re-commission.

Where it breaks: don’t commission this where the answer lives behind a paywall, inside your own systems, or in a story still unfolding this week. The web isn’t the source there, and a polished cited report will hide that it guessed.

Additional reading

• Deep Research Max: a step change for autonomous research agents — Google (Apr 2026) — the source for “extended test-time compute” and the asynchronous “overnight due diligence” background-run framing.
• Exploring Deep Research: Three Tips for Better AI-Assisted Inquiry — OpenAI Forum (Apr 2025) — define the objective, iterate the brief, and avoid both over- and under-instructing.
• Deep research in ChatGPT (FAQ) — OpenAI Help Center — how the agent clarifies intent before running, consults many sources, and why a specific brief beats a broad one.

Editor’s note

There’s no absolute rule for how to write a brief for deep research, or for any large job you hand to an agent. Extensive plan iteration over many turns is my usual method — I’ll often spend an entire session (or more) planning, then hand the resulting brief to a new session when I’m ready. You don’t have to do it that way; depending on how you manage your context, another method may produce better outcomes for you. The critical starting point you need to reach is a well-developed brief — get there, and you’re in good shape for the long-running task.