Deep Research Agent

A deep research agent is an agent that plans a multi-step investigation, issues many searches, reads the sources it finds, and returns a cited synthesis rather than a single answer. Where a search API call returns a ranked set of results and a standard RAG pipeline returns one grounded response, a deep research agent runs a loop: decompose the question, fan out across sub-queries, read what comes back, notice what is still unanswered, search again, and stop when its own sufficiency criteria are met.

Four mechanics define the category. Planning turns the prompt into sub-questions, which is where most of the quality difference between systems lives. Fan-out issues those sub-queries, often dozens of them and often concurrently. Iterative refinement feeds results back into planning so the second round of queries is informed by the first. Citation assembly maps each claim in the output back to a source URL, which is what makes the report checkable at all; without it a reviewer has to re-run the research to verify a single line.

Two costs are routinely underestimated at purchase. The first is latency measured in minutes. OpenAI's own launch post for deep research put a run at five to thirty minutes. Perplexity's announcement of its Deep Research agent said most tasks finish in under three minutes. Parallel's published comparison table lists its own Task API tiers running from ten to sixty seconds at the Lite tier up to minutes or hours at Ultra8x. None of that fits in a request-response path where a user is waiting. The second is token spend, which scales with depth rather than with question count. Parallel publishes prices for its own tiers ranging from $5 per thousand queries at Lite to $2,400 per thousand at Ultra8x, alongside BrowseComp accuracy of 4 percent and 58 percent respectively. That is a vendor's own comparison and the competitor rows in it deserve the usual skepticism, but the internal spread is the point: within one product line, buying more accuracy costs several hundred times more per query.

For a product builder this decides the architecture before it decides the vendor. Deep research belongs in an asynchronous job with a notification on completion, not behind a chat cursor, and it needs a spend ceiling per run because the loop decides its own length. It also needs deduplication, since the same authoritative source will surface across many sub-queries and paying to read it repeatedly is waste. Pick the accuracy target up front: the tier decision is a direct trade of dollars against correctness, not a preference.

As of mid-2026 the pattern across vendors is a research endpoint layered above a search endpoint: Exa, Tavily, and Parallel each expose both, with the research tier priced per run rather than per query, and Perplexity's Sonar family sells the answer-synthesis step as its product. Evaluate them on your own questions under a fixed budget, and check how many of the returned citations actually open and support the claim they are attached to.