The online research skill — a documented, repeatable scenario the agent follows step by step — is Research Agent’s main use case: a genuinely deep study of any topic. Not one query with a couple of sources, but dozens of searches, downloading and analysis of hundreds of pages, batch data extraction and a verified report in the research library.
A fixed pipeline
The research runs as a fixed linear pipeline, with no improvising along the way:
- create session — a named research session that groups every operation of the topic;
- search — searching in two phases: first two broad scouting queries on the topic (one Russian, one English) with a read-through of the results, then eight more targeted queries built from the subtopics, sources and terms the first round surfaced — so the bulk of the searching lands on target;
- fetch — filtering the results for relevance and downloading every suitable page;
- extract — chunking the downloaded Markdown (splitting it into pieces the AI model can digest) and batch extraction of structured data;
- write — writing the report as Markdown with frontmatter (a small header of title and order) into an existing category of the research tree;
- verify — verification: searching for sources that confirm and sources that refute.
Waiting at the agent’s pace
Every operation is asynchronous: the start call returns an id, and the agent then polls the status on the interval the server itself names (the expected delivery time). Sessions show the statuses Pending, Running, Completed and Failed, plus the confidence of the estimate. Search, fetching and especially LLM extraction (having the AI model pull structured data out of the pages) calmly take minutes — large extractions take more than ten. That is normal progress, not a hang-up.
The report
The report is written in Markdown with frontmatter (title, order) and appears in the research tree immediately — the library refreshes in real time. Before saving, the text is normalized through markdown/clean. Each report lands in an existing category of the tree: new categories are never created without the owner’s request. The agent does not keep asking the owner: when something could not be collected, it continues with what it has, and describes the lost sources and the decisions it made separately in the report.
Verification
The finished report is not passed off as the ultimate truth: the agent looks for sources that confirm and sources that refute the key claims, and adds a critique section with alternative interpretations to the report. The primary sources (search sessions and downloaded pages) stay attached to the report — any fact can be checked.
Completeness over brevity
The skill’s main rule: the report’s goal is completeness, not brevity. A table of 50 metrics stays a table of 50 metrics, a list of 30 tools stays a list of 30 tools. Facts, numbers and sources are never thrown away for the sake of tidiness: every dropped number is a decision made blind.
Economics
In our own runs, one research on the DeepSeek-V4-Flash model costs 9–27 cents and finishes in 20–30 minutes — at a depth that services like Perplexity or ChatGPT with search cannot reach. Token and cost statistics for every operation are available in the Completions section.