Claude Code's new cross-session messaging enables direct agent-to-agent communication
Anthropic
Anthropic has released Claude Code v2.1.224, introducing cross-session SendMessage, allowing separate Claude Code sessions to discover each other via ListAgents and communicate directly over Unix sockets. A user describes how this feature enables a coordinator-satellite setup, reducing token usage compared to a single universal agent, and notes that v2.1.225 extends this to remote machines with remote control enabled.
An AI user reports that in the latest Claude Code version 2.1.224, two agents working on the same problem unexpectedly exchanged messages directly through a Unix socket, a feature never seen before. This new cross-session SendMessage lets different running Claude Code sessions find each other via ListAgents and send messages to each other over sockets like recipient: uds:/tmp/cc-socks/769334.sock. Previously, such interactions required external channels, which was so inconvenient that the user ended up using a single universal agent for 90% of work, consuming a large context and burning twice as many tokens as a multi-agent scheme. The user initially wanted a system with many agents working on different projects, communicating what they do, with a main coordinator. Today's v2.1.225 release allows Claude Code instances on different machines with remote control enabled to exchange messages, which was not possible yesterday. This opens up the possibility of creating a Claude coordinator that launches or connects to full Claude Code instances in projects and distributes tasks to multiple independently running Claude Code instances, collecting results and sending new tasks, a level above regular subagents.
Source: Habr — хаб ИИ —
original
