Plan, fan out subagents with separate context windows, verify every claim against a source, then synthesise — the shape behind every 'deep research' product.
Sending four researchers to different libraries with clear briefs, then writing the report from their notes rather than from every book they touched.
Breadth costs context, and one window can't hold five lines of inquiry. Separate windows per subagent is the mechanism, not the parallelism.
A research question is not a retrieval question: it needs breadth, comparison, and a defensible answer. The pattern that works is an orchestrator with subagents. The lead agent decomposes the question into independent lines of inquiry and spawns a subagent per line, each with its OWN context window — which is the real point, since parallel search would otherwise fill one window with hundreds of pages of raw source. Subagents search, read, and return compressed findings with source ids. The lead synthesises, and a separate citation pass attaches every claim to a source, dropping claims that cannot be supported. Anthropic reported that this multi-agent shape substantially outperformed a single agent on breadth-first research, and that token spend — roughly an order of magnitude over chat — is what buys the difference.
A deep research agent decomposes a question into independent lines of inquiry and runs a subagent on each, every subagent with its own context window so raw sources never crowd the orchestrator's. Subagents return compressed findings with source ids; the lead synthesises; and a separate citation pass binds each claim to a source and drops what can't be supported. The context isolation and the separate citation pass are what make it work.
Deep Research Max: A step change for autonomous research agents — Google for Developers, 3:03