Claude Code agents can now talk to each other mid-task

If you use Claude Code across multiple terminal windows, you know the drill: session A finishes a database migration, and you have to manually copy the schema over to session B and explain what changed.

Anthropic is putting an end to that manual context shuttling. Claude Code now supports cross-session messaging, allowing parallel AI coding sessions to talk directly to each other on your machine and across the cloud.

Slack-style mentions for your terminal

Under the hood, Claude Code uses two new tools, ListAgents and SendMessage, to discover and route messages between sessions. You don't have to call either of them directly.

Starting with v2.1.232, the interface works just like tagging a colleague. You type @ followed by the first letters of another session's name to pick it from a typeahead list, like @api-worker. Names with spaces or special characters can be wrapped in quotes, such as @"release notes". If multiple active sessions share a name, Claude stops to ask which one you meant.

Once tagged, Claude generates a summary of whatever finding or decision needs to move over. But it doesn't always wait for you to prompt it: Claude can autonomously message another session when it discovers a breaking change that impacts parallel work, or when coordinating code landing across parallel Git worktrees.

It also handles long-running jobs. You can instruct a test suite or migration running in another session to ping your active terminal with an update the moment it finishes or goes idle.

Safe handoffs and strict boundaries

The delivery mechanism is built to avoid breaking active work. Incoming messages arrive as plain text and wait between tool executions, meaning a running script or compile command is never interrupted mid-flight. If the target session is sitting idle, the message simply wakes it up to start a new turn.

Crucially, permission boundaries do not transfer. Each session stays locked to its own security perimeter. Anthropic explicitly instructs Claude never to ask another session to perform an action that was denied or blocked in its own environment.

Claude Code also enforces hard guardrails: messages exceeding a size cap are killed before leaving, rapid bursts that overwhelm a target inbox get cut off, and unsafe reply targets like symlinked directories are blocked outright. Incoming messages can be delivered, held for manual approval, or refused.

The feature works out of the box on macOS and Linux (including WSL 2) running Claude Code v2.1.224 or later, and on native Windows starting at v2.1.234. It even bridges across different physical machines and Remote Control cloud sessions.

Why it matters

For power users running multi-agent workflows, this is the missing link. Human copy-pasting between terminal tabs has been the single biggest bottleneck in running parallel AI coders. Being able to kick off a migration in one worktree and have Claude alert your API session when it lands is an immediate productivity win, and it is ready to use today without extra configuration.

The catch? Every delivered cross-session message counts toward your token usage just like a user-typed prompt. If sessions start chatting back and forth autonomously, your API bill could climb quickly.

Strategically, we think Anthropic's approach is brilliant. Instead of forcing developers into a proprietary multi-agent GUI or a rigid orchestration platform, they simply turned the standard CLI into a networked mesh of agents. Keeping strict permission boundaries per session is smart engineering that stops prompt-injection escalation across workspaces. The real question is how trigger-happy Claude will be with autonomous messages, and whether developers will find themselves paying for conversational fluff between their own terminals.