Deep research agents

Why it matters

A single search and an answer is a lookup. A research question, the trade-offs between two approaches, what changed in a regulation, which suppliers of a part are credible, needs many searches, many pages, judgement about which sources to trust and a written result someone can check. Deep research agents run that loop for minutes and return a report with citations. Their cost is dominated by reading rather than acting, and their worst failure is a confident report built on thin evidence. Interviewers ask about them to test whether you can bound an open-ended loop and evaluate prose that cites.

The idea

The loop

Plan: turn the question into sub-questions, each answerable by search. Search: run several queries per sub-question through a search tool that returns titles, snippets and URLs. Read: fetch the pages worth reading, as extracted text rather than raw HTML, and skim for the sub-question. Extract: pull claims with their source into a structured note: claim, source, date, confidence. Gap check: compare the notes against the plan; which sub-questions are thin, which claims rest on a single source, which contradict each other. Search again on the gaps. Synthesise when the stop condition fires: a report argued from the notes, with a citation on every claim that came from a source.

Source quality and deduplication

The web returns the same press release on forty domains and the same claim copied between blogs. Deduplicate by canonical URL and by near-identical text, then weight sources: primary (the standard, the filing, the paper, the vendor's own documentation) over secondary over aggregators; recent over stale for anything that changes; and never a single source for a number. The agent keeps a per-source note on why it trusted it, so the report can say "according to the regulator's notice" rather than "sources say".

Citation discipline

A citation is a link from a claim to a passage, kept from extraction through synthesis. The report is written from the notes, not from memory of the pages, and each sentence that asserts a fact carries the note's id. Claims with no note are marked as the agent's inference. That discipline is what makes the report checkable, and it is what breaks first when the synthesis step is handed the raw pages instead of the notes: the model writes fluently and the citations become decoration.

Stop conditions and cost

The loop must end. Two conditions: coverage, every sub-question has at least two independent sources or an explicit "nothing found"; and budget, a cap on searches, pages read and tokens. Reading is the cost. A page is a few thousand tokens, a run reads thirty to a hundred, and only the notes, not the pages, enter the synthesis context. Sub-questions are independent, so they run in parallel, each as a worker with its own context, which cuts wall-clock time from many minutes to a few and keeps every context small. The Orchestrator-workers lesson covers that shape.

Evaluation

Two measures. Factuality against a rubric: for a question with known answers, does the report contain the key facts, and does it contain false ones. Citation precision: sample cited claims, open the source, check that it supports the claim. Add coverage (sub-questions answered) and cost per report. Human review of a sample stays necessary, because a fluent wrong report is the thing no automatic grader catches reliably.

	Single search plus answer	Multi-query RAG	Deep research agent
Depth	One query, top results, one answer	Several rewritten queries over an index, one synthesis	Sub-questions, iterative search, cross-checking, gap filling
Latency	Seconds	Seconds to tens of seconds	Minutes
Cost	A few thousand tokens	Tens of thousands	Hundreds of thousands to millions; reading dominates
Citation quality	Links to the top results, loosely tied to claims	Passage-level citations from the index	Claim-level citations from notes, with source tiers and dates