Source: Anthropic (Claude Code docs) — 2026-08-07

Summary

Anthropic shipped self-hosted environments in public beta for Team and Enterprise Claude Code plans: running claude self-hosted-runner setup turns an organization's own machines or containers into runners, so cloud sessions launched from claude.ai, mobile, desktop, or claude --cloud execute inside the customer's own network, keeping repo checkouts, build artifacts, and secrets off Anthropic's infrastructure. The same week 32 release (v2.1.220-v2.1.224) added cross-session messaging — a ListAgents tool to discover other running sessions and a SendMessage tool to send them text, either on request or autonomously. The release also flips "auto" to the default permission mode for new sessions starting August 14, removes the 200-subagent-per-session cap, and tightens worktree isolation to block Bash/git commands that reach outside the isolated worktree into the main checkout.

Key Takeaways

• Self-hosted environments change the trust boundary, not the product: the same cloud-triggered sessions now run on infrastructure the customer controls, so secrets and proprietary code never have to leave the org's network.
• This directly targets the objection enterprises raise about cloud coding agents — "where does our code actually execute" — by letting Team/Enterprise plans answer "on our own machines."
• Cross-session messaging via ListAgents and SendMessage turns isolated coding sessions into a coordinating swarm: the docs' own example is one session autonomously telling another "the users.name field changed" — agents flagging breaking changes to each other without a human relaying the message.
• Removing the 200-subagent cap signals Anthropic expects genuinely large multi-agent fan-out workloads now, not just occasional parallelism.
• "Auto" becoming the default permission mode (from August 14) is a meaningful default-behavior shift — new sessions will act with less per-step confirmation unless a team explicitly dials permissions back.
• Tighter worktree isolation blocking Bash/git commands from reaching outside the sandboxed worktree closes a real escape hatch — it's a direct hardening response to agents being able to touch the main checkout from an isolated branch.

Reel Script

Hook (18s)

Claude Code agents can now message each other mid-task — one session finishing a database migration can autonomously warn a completely different session that a field it depends on just changed. No human in the loop.

Core Concept (70s)

Two separate things shipped together here, and both matter. First: self-hosted environments. Normally when you kick off a Claude Code session from the cloud — your phone, the web, wherever — it runs on Anthropic's servers. Now, Team and Enterprise orgs can run a setup command that turns their own machines into runners, so that same cloud-triggered session executes inside their own network instead. Your repo, your secrets, your build artifacts never leave your infrastructure — think of it like choosing to host your own mail server instead of trusting a third party with your inbox. Second, and more interesting: cross-session messaging. Claude Code sessions can now discover each other with a tool called ListAgents, and send each other plain text with a tool called SendMessage. This can happen because you told it to, or the agent can just decide to do it on its own — Anthropic's own example is one session noticing a breaking API change and proactively telling the session working on a different part of the codebase.

Hands-On (60s)

Picture the actual flow: Session A is refactoring a users table and renames a field. Instead of that just silently shipping and breaking something downstream, Session A can call ListAgents, see that Session B is active and working on the payments API, and call SendMessage to say, in effect, "the users.name field changed, you may need to update your code." That's inter-agent communication happening peer-to-peer, without a human relaying context between two chat windows. Same release window, three other changes worth knowing: "auto" permission mode becomes the default for new sessions starting August 14, meaning less manual approval friction by default; the old 200-subagent ceiling per session is gone entirely; and worktree isolation got tightened so a session sandboxed to one git worktree can no longer run Bash or git commands that reach out into your main checkout.

Takeaway (25s)

My take: cross-session messaging is the bigger deal of the two — it's the first real building block for agents coordinating on a shared codebase without a human as the message bus. If your team runs multiple Claude Code sessions in parallel, go read the week 32 changelog before August 14, because your default permission behavior is about to change.

Discussion