Every deep-research session on video content has the same three movements: find the sources worth reading, capture them somewhere stable, then interrogate them until the question is answered. Done manually — even with good tools — you're the conductor: run the search, review, trigger the imports, wait, ask the questions, repeat for the follow-ups.

The conductor role is delegable. MCP agents chain tools natively: one well-shaped instruction runs the whole pipeline — search, curate, import, interrogate — and returns with the cited answer, sources archived. You state the research goal; the movements happen.

The chain is three tools your assistant sequences on its own. deep_search finds and ranks the candidates — quality verdicts doing the curation your review pass used to. library_import_youtube_video captures the top sources into your library, permanent and re-queryable. deep_context interrogates the imported set, returning answers cited to the chunks.

One prompt describes the whole arc: "research <question>: find the strongest sources, import the best ones, then answer from them with citations." The agent runs the movements, you receive the answer plus a durable asset — the imported sources, which every follow-up question queries for the cost of the question. It's the pattern underneath half the library-based use cases, stated once and reusable for any topic.