Claude Code agents can now message each other mid-task

Claude Code sessions can now send messages to each other, eliminating the copy-paste shuffle developers run when multiple AI agents work on related tasks. Anthropic's update, documented today, lets one terminal ask another session a direct question and receive the answer inline. Sessions can also fire off unsolicited alerts, so an agent that changes a shared module can warn a sibling session before it trips over stale context.

How inter-session messaging works

On a single macOS or Linux machine, communication stays local. One session sends a text summary or query, the recipient processes it, and the response comes back without touching Anthropic's infrastructure. The feature also works across different computers, but in that scenario messages route through Anthropic's servers and only responses are supported. That asymmetry matters: you can pull information from a remote session, but you cannot push to it proactively.

Anthropic's documentation lists three typical use cases: sharing insights discovered mid-task, coordinating parallel workflows, and checking status on long-running operations. The last one is the most practical. Instead of tabbing to another terminal and manually asking "are you done yet?", one agent can poll another and act on the answer.

Platform limits and admin controls

The feature ships only for macOS and Linux. Windows users get nothing, and neither do teams running Claude Code through Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry. Those managed environments lock out the messaging layer entirely.

For enterprises on supported platforms, admins can disable inter-session communication via settings. That matters for compliance-heavy teams worried about agents coordinating in ways that are hard to audit. The control exists, but Anthropic hasn't published detailed logging or observability hooks yet.

What this changes for multi-agent workflows

Until now, running multiple Claude Code sessions meant manual context transfer. You'd copy a summary from one terminal, paste it into another, and hope the second agent understood the context. This update removes that friction for tasks that split naturally, like running a backend refactor in one session while a second session updates the corresponding API client.

The coordination also opens up a supervision pattern. One session can act as an orchestrator, dispatching subtasks to worker sessions and aggregating their results. Anthropic's documentation doesn't prescribe this architecture, but the primitives are now in place.

What Anthropic didn't say

The documentation is silent on rate limits. If one session can query another, can it spam it? There's no mention of throttling or priority queuing. Latency guarantees are also absent. For local communication the overhead should be negligible, but cross-machine messages routed through Anthropic's servers could introduce delays that break time-sensitive coordination.

Token consumption is the other open question. When one session sends a summary to another, does the receiving session pay for those tokens as fresh input? Probably yes, but Anthropic hasn't clarified billing implications for heavy inter-session chatter.

Source: The Decoder / Matthias Bastian