Claude may get multi-agent Research Mode with memory and task delegation

Anthropic continues to refine its research mode feature, which has been renamed from its previously used internal name, Compass. Recently, more details about this feature have surfaced. It appears that the Claude Research feature may actually support multi-agent functionality. This would include the use of various tools such as web search, thinking, and sleeping, along with new capabilities that haven’t been present before, like memory and a tool called “create sub-agent.”

The “create sub-agent” tool potentially allows a master agent to spawn tasks for sub-agents. Within a single research flow, users could witness actual multi-agent interaction, making the entire process more dynamic and collaborative.

Research tools that will be available to Claude's Agent (discovered by Tibor Blaho):

• web_fetch - crawling an individual URL
• web_search - using Brave web search
• create_subagent - potentially delegating a task to another sub-agent
• memory - recalling to its own memory
• think - thinking (existing feature of Claude Sonnet 3.7 Thinking model)
• sleep - unknown
• complete_task - potentially preparing a summary response
• batch - unknown

This positions Anthropic in a unique spot compared to similar tools available in other apps, such as Deep Research in Gemini or the research features in ChatGPT, where the models typically focus on gathering information from a large number of internet sources to produce a summary.

Recently, Anthropic also revamped its UI as a part of the preparation to upcoming feature releases as well. It will be interesting to see how this evolves. While the feature still seems far from being fully ready, the possibilities it suggests are quite exciting.