Deep Research

This is a lead-agent-plus-subagents architecture for writing research reports with evidence tracking and citation governance. The lead agent dispatches parallel research tasks to subagents who write distilled notes, then builds from those notes instead of raw search results, which cuts context usage by 60-70%. You get source type labels (official, academic, journalism, community), freshness checks against an AS_OF date, and counter-review passes before final output. It enforces quality gates like minimum 30% official sources and maximum 25% from any single domain. The architecture splits into enterprise mode for company research with SWOT and risk matrices, or general mode for industry and policy work. Useful when you need a proper literature review or competitive analysis instead of a quick summary.