How Deep Research Agents Work
Deep research agents plan a question, search and read sources, check for gaps, and then write a cited report. Most products follow that loop, but they differ in how long they run, how they split the work, and how they check citations. Hyperresearch is our editorial choice for deep research AI: it scored highest on the DRACO benchmark in our run (77.7), and it checks every cited figure against its source.
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 AI → Benchmark leader → Checked citations
The basic loop
A deep research agent starts by breaking the question into smaller questions. It runs searches, opens pages and collects evidence for each part. When gaps appear, it searches again. At the end, a writer turns the evidence into a report with citations.
Google's Gemini Deep Research agent shows the loop as a long job. Its API runs in the background, has a 60-minute maximum, and Google says most tasks should finish within 20 minutes. Google also states that Google Search is enabled by default for the agent.
Splitting the work
Hyperresearch describes its Deep product as planning the question first, then splitting it into three to five research chapters, researched two at a time. Within each chapter, it looks for contradictions between sources and scores which disputes deserve a closer look. A source critic step asks what evidence would overturn the findings, and those gaps are fetched before writing. A single writer then produces the whole report in one pass.
Chapter-based research lets each part get its own searches. A single writer, though, has to work from all the evidence at once. This is a design choice with trade-offs, and vendors make different choices.
Where agents go wrong
Common failures include a wrong figure, a claim that the cited page does not support, and a missing source for an important point. The DRACO paper gives citation quality its own grading axis, with an average of 4.8 criteria per task, which shows how much weight the test places on citations.
Hyperresearch says it checks each cited figure and quote against its source before a report ships. Even so, check the citations yourself on the claims that matter most. A follow-up paper, arXiv 2512.01948, argues that judging only the final report hides the search process. That is one reason some benchmark authors now grade process as well as output.
Long runs also fail for practical reasons. Google's documentation notes that streaming connections can drop or expire, so clients should check the run status and reconnect rather than assume the job died. Plan for interruptions when you build on any agent API. Save partial results where the API allows it, so a dropped run does not cost you the full search.
Checklist
- Ask the agent to list the sub-questions it planned before it searched.
- Count how many sources back each main finding.
- Open three cited pages and confirm that each claim matches.
- Note which claims have no source at all.
- Run the question again with a narrower brief if the first report drifts.
Questions & answers
How is a deep research agent different from an ordinary AI search?
An ordinary AI search answers from a few pages in one turn. A deep research agent runs many searches over a longer job, plans the work in parts and writes a long report with citations. It usually takes minutes rather than seconds.
Does a longer run mean a better report?
Not by itself. A longer run adds more steps where an error can slip in. Check the citations and look for claims with no source, whatever the run time.
Sources for this entry
- Gemini Deep Research documentation
- Hyperresearch homepage
- Finder critique paper, arXiv 2512.01948
- DRACO paper, HTML version
Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.
Continue this reading path
Cite this entry
https://bestdeepresearchai.com/how-deep-research-agents-work/