Research

Ask a normal question and Claude answers it. Turn on Research and something different happens: Claude plans an investigation, spawns parallel agents to search simultaneously, reads what comes back, decides what it still does not know, and searches again — until it can write a cited report.

Why it is worth understanding architecturally

Research is the most visible consumer-facing use of the same idea as subagents in Claude Code: a lead agent delegating to parallel workers, each burning its own context on a slice of the problem and returning conclusions rather than raw material.

Anthropic’s published engineering account reports that a multi-agent system with Opus 4 leading and Sonnet 4 subagents outperformed single-agent Opus 4 by 90.2% on their internal research evaluation. That number is from that specific evaluation on that model generation — treat it as evidence the architecture matters, not as a figure to quote about current models.

What it can search

The web, Google Workspace, and any connectors you have enabled. With Workspace connected it can search email, review documents and read calendar commitments — which is exactly the capability worth thinking about before enabling it on a work account.

What it does not do

• Not on the Free plan.
• Not on the API. Research is a Claude apps mode. To build something comparable you assemble it from tool use, the web search server tool and your own orchestration.
• Not fast. It is doing many searches; that is the point.
• Not immune to bad sources. It cites what it found. Citation is provenance, not verification.