Create Blog Article “2026-09-21_this-weeks-ai-bite-claude-code-sessions-can-now-talk-directly-to-one-another”

content/posts/2026-09-21_this-weeks-ai-bite-claude-code-sessions-can-now-talk-directly-to-one-another.md

@@ -0,0 +1,49 @@
+---
+title: "This week's AI Bite: Claude Code sessions can now talk directly to one
+  another"
+cover: https://res.cloudinary.com/de4rvmslk/image/upload/v1790000557/Claude_Code_Karol_hnogzo.png
+tags:
+  - Claude Code
+  - AI Agents
+  - Multi-agent Systems
+  - Developer Tools
+  - AI Coding Assistant
+  - Software Engineering
+postAuthor: Karol Kasprzak
+---
+As a Senior Software Engineer, I stumbled onto something by accident this week – a capability that's apparently been quietly available in Claude Code for a month. Two separate sessions of mine were able to exchange messages with each other.
+
+**The mechanism**
+
+Two commands make this possible. ListAgents lets a session see which other Claude Code sessions are currently running on the same machine. SendMessage delivers a message to a chosen session, functioning exactly as if a person had typed it directly into that session's terminal – and the receiving session responds the same way.
+
+**How I used to handle this**
+
+When running several agents in parallel – each working in its own worktree on related pieces of a task – getting information from one to another meant either manually copying text between terminal windows, or rigging up a makeshift relay: a shared file where one agent would write and another would check for updates. Functional, but clunky.
+
+**What happened this time**
+
+![](/img/claude-code-session.png "Claude code session")
+
+A reviewer raised a question about an implementation detail that only one of my sessions (Session A) had the context for. Rather than switching terminals and re-tracing the code myself, I simply asked Session B to relay the question. It came back with a full answer, reasoning included, which I passed straight to the reviewer – no manual copying involved.
+
+**Not the same as subagents**
+
+The first reaction I got when sharing this was: "isn't that just how the main agent already talks to its subagents?" Not quite. What's happening here is two fully independent Claude Code instances communicating – separate processes, separate worktrees, separate contexts entirely. And it's not confined to a single machine either: with Remote Control enabled, sessions running on different computers can reach each other too.
+
+**A few practical details**
+
+* Available since v2.1.224 (August 7) on macOS, Linux, and WSL 2; native Windows support arrived in v2.1.234.
+* No setup required — it's active by default.
+* Only plain text is exchanged; neither conversation history nor files are shared automatically. To transfer full context, you still need /resume.
+* As of v2.1.232, tagging a session with @session-name in your prompt is enough to send it a message.
+* Sessions can message each other autonomously, without a human initiating it, if you allow that. This is governed by the crossSessionInbound setting (accept / hold / refuse). Messages sent off the local machine can also be configured to require your manual approval, even under bypassPermissions.
+
+
+**Things to keep in mind**
+
+* The feature isn't without growing pains: after an automatic update from 2.1.258 to 2.1.260, some users found SendMessage had stopped working in sessions that were already open. If something seems broken post-update, this is a known issue.
+* One question worth sitting with: if agents can talk to each other unsupervised, could one end up interfering in a session it has no business touching? The honest answer is yes — but only if you've configured it to allow that. That's precisely what crossSessionInbound: refuse is there to prevent.
+
+
+For anyone juggling multiple agents on interconnected tasks, the manual back-and-forth just became optional – at least in my workflow.