Deep Research

This orchestrates multi-phase research campaigns that pull from web search, codebase analysis, and stored memory to answer complex questions. It breaks topics into sub-questions, searches existing knowledge first, gathers new information from web and code, then cross-references everything into a structured report with evidence quality ratings. You get four depth levels from quick (2-3 minutes) to exhaustive (30+ minutes with parallel sub-agents). The useful bit is that it stores findings in organized memory namespaces as it goes, so repeated research builds on itself rather than starting from scratch. Good for investigating architectural decisions, debugging obscure issues, or any question that spans multiple sources and needs more than a single search query.