Claude Code Adds Cross-Session Agent Messaging
16 Aug 2026
Claude Code has introduced a cross-session messaging feature that lets separate agent sessions discover and communicate with one another, according to the announcement. Sessions can now use ListAgents to find reachable agents and SendMessage to deliver messages between them.
What's new
Claude can message subagents and agent-team teammates within a single session or team. A new command, /list-agents, lets users see which sessions Claude is able to reach. Interactive sessions are given default names derived from their working directory's folder — for example, a folder called "myapp" might produce a session name like myapp-3f. Notably, a claude -p session also binds an inbox socket like an interactive session and can receive messages.
How messages are handled
Every incoming message is checked against the receiving session's own inbound controls, producing one of three outcomes:
- Delivered
- Held
- Refused
The report does not specify what conditions trigger "Held" versus "Refused" — this distinction remains unclear.
The system includes several deliberate safety boundaries. Messages arriving from another session cannot approve permissions, change configuration, or execute commands. There's also a default dialogExpiry window of five minutes for message approval, and Claude cannot initiate a conversation with sessions on other machines — it can only reply to messages that originated from them.
Why the limits matter
Message delivery is not guaranteed in every configuration, meaning coordination signals between sessions could occasionally be missed. The restrictions on cross-machine initiation and message capabilities point to intentional — if limiting — safety design. Reliance on default folder-derived naming could also introduce ambiguity in complex, multi-session setups where many sessions might have similar or overlapping names.
Potential use cases
According to the report, this feature could open up several coordination patterns for engineering teams:
- One session handing off findings to another, potentially cutting down on manual re-explanation of context.
- Sessions working on the same repository across separate git worktrees coordinating about what changes have landed.
- Long-running tasks — like migrations or test runs — reporting status back to a session that's being actively watched, which could streamline monitoring workflows.
Why founders should care
For founders building with or on top of Claude Code, this feature likely signals a broader shift toward multi-agent orchestration in AI-assisted development. If your team runs parallel Claude Code sessions — across worktrees, subagents, or long-running jobs — this could plausibly reduce coordination overhead by letting sessions communicate directly rather than requiring a human to relay context between them.
At the same time, the explicit safety restrictions (no permission approval, no command execution, a five-minute approval window) suggest Anthropic is deliberately keeping human oversight in the loop rather than enabling fully autonomous multi-agent handoffs. Founders evaluating this for production workflows should weigh the current gaps: there's no confirmed information yet on release timing, pricing or plan-tier availability, performance/scale limits, or the security model for cross-machine communication beyond the reply-only restriction. Teams considering early adoption may want to treat this as an experimental coordination layer rather than a fully hardened orchestration system for now.
What's still unclear
Several open questions remain, including the exact criteria separating "Held" from "Refused" messages, the rollout timeline, pricing or tier gating, latency and scale characteristics at higher session counts, and further security details for cross-machine scenarios. Founders planning to build workflows around this feature should watch for follow-up documentation on these points before committing to it for critical paths.