Deep research by AI agents: how it works and how to review it

Ask an assistant a question and it answers in one pass. Ask a deep research tool the same question and it goes away for a while, runs search after search, reads what it finds and comes back with a long report full of footnotes. The second looks far more authoritative, which is exactly why it needs a different kind of review. This page explains what deep research is, what happens at each stage, which questions it suits, and how a team can check a report before anyone acts on it.

What deep research means

Deep research is multi-step research carried out by an AI agent. Given a question, the agent plans an approach, gathers sources over many rounds of searching and reading, follows leads it did not anticipate, reconciles what it found and writes a structured report in which claims point to sources. The defining feature is the loop: what the agent reads in round three changes what it looks for in round four.

A useful mental model is a quick, tireless junior researcher: excellent at assembling material, unreliable at knowing which material deserves trust.

The four stages, and what goes wrong in each

1. Planning

The agent breaks the question into sub-questions and decides where to look. A good plan is visible before the work starts, so you can catch a misreading early. A common failure is quiet reinterpretation: asked about regulatory risk for a product in Germany, the agent researches EU regulation in general, because that is where it found the most material.

2. Gathering

The agent searches, reads, pulls out relevant passages and follows references. Failures here are about sources: leaning on aggregator sites and press releases instead of primary documents, missing paywalled or non-English sources entirely, and stopping when results start to repeat rather than when the question is answered.

3. Synthesis

The agent combines its findings into a narrative. This is where disagreements between sources should surface, and where they often vanish instead: two conflicting figures become one, or a minority view drops out because it appeared only once.

4. Citation

Each claim is linked to a source. The risk is a citation that points to a real page which does not actually support the sentence attached to it. A footnote shows that a page was visited, not that it says what the report claims.

When deep research helps, and when it does not

It earns its keep on questions that are broad, answerable from accessible sources and worth a structured first pass:

• A landscape scan of companies, products or approaches in a market.
• A first-pass literature review before an expert reads the key papers.
• Compiling how a rule or standard is described across several jurisdictions or bodies.
• Background on a company ahead of a meeting or a due diligence process.

It is a poor fit when the answer lives where the agent cannot reach (internal systems, paid databases, conversations), when you need one specific fact (a plain search is quicker and easier to check), when the facts change faster than the report can be reviewed, or when the real work is a judgement call. A report can inform a decision; it should not be mistaken for one. If you are weighing whether to build this capability or buy it, build a research agent or buy deep research walks through the choice. Choosing a deep research API covers the tools, and research agents that hold up is for teams building their own.

Brief it like a researcher, not a search box

Most weak reports trace back to a vague request. A useful brief fixes the question, the boundaries and what done looks like before the agent starts:

Question: Which payment providers offer instant payouts to marketplace sellers in Turkey, and under what conditions? Scope: providers with a publicly documented offering for marketplaces; leave out announcements of future products. Sources: the providers’ own documentation and official regulator pages first; news only to corroborate. Output: one row per provider with the condition, the page it comes from and the retrieval date. Done when: every provider found has a row, or is listed as “no public information found”. If sources disagree, show both.

Two parts of that brief carry most of the weight. The source rule tells the agent which evidence is allowed to answer, a choice worth making deliberately, as deciding which sources may answer explains. And the “done when” line turns gaps into something to report instead of something to fill.

If the question changes halfway through, write a new brief. Letting the agent drift toward a broader question produces a report that answers neither version well; a changed question needs a new plan makes the case.

How to review a deep research report

Do not read the report from start to finish and judge whether it sounds right. Check it.

1. Compare the question. Does the report answer the question in the brief, or a broader one?
2. Sample the citations. Pick a handful of the claims that matter most and open their sources. Does each passage say what the sentence says, with the same strength?
3. Look at the source mix. How much rests on primary documents, and how much on summaries of summaries?
4. Hunt for missing disagreement. Search on your own for the strongest counter-evidence to one key claim. If it exists and the report ignores it, treat the synthesis with caution.
5. Check the dates. Are the sources recent enough for this question?
6. Read the gaps. A report that found an answer to everything is suspicious. Good ones list what they could not establish.

Citations that hold up goes deeper on the second step.

Deep research with a plan you approve

GoCux treats research as a plan before it treats it as a report. A plan holds the question, the ordered steps and an allowlist of sources. Creating it records a draft, and it runs only after someone approves it; while running, it cannot go beyond what was approved, and an allowlist left empty permits no source at all. If the question changes, the right move is a revised plan rather than a quiet expansion. Findings keep two kinds of result apart: a finding backed by a quotation, and a question the sources could not answer. Both are retained, so the gaps in a report stay visible instead of being smoothed over. A citation remains a pointer to evidence rather than the evidence itself, and the reviewer still checks that the passage means what the finding claims. The Task page shows how work like this moves through visible states with a decision point, and the source policy documentation explains allowlists.

Checklist before you rely on a report

• The brief states the question, scope, source rules, output format and when the work is done.
• The plan was visible, and approved, before the research ran.
• The report answers the question as written, not a widened version of it.
• You opened the sources behind the claims that matter most, and they say what the report says.
• Conflicting sources are shown, not averaged away.
• Questions without a source are listed as open, not filled in.
• Someone with domain knowledge has read the report before it informs a decision.