Claude Code rolled out version 2.1.211 on July 15, adding a “--forward-subagent-text” flag and a bundle of stability and security patches aimed at multi-agent workflows. Teams that run Claude Code as a backend service can finally watch subagents in real time and avoid the silent failures that have plagued long-running jobs.

Why the update matters now

Claude Code’s core value lies in orchestrating multiple AI agents that hand off work to one another. Until now, a parent task could appear idle while a subagent was busy, and agents sometimes resurfaced with stale prompts after being killed. Version 2.1.211 fixes that by giving you visibility, reliable state handling, and tighter permission gating.

The new subagent-visibility flag

The “--forward-subagent-text” option streams a subagent’s textual output and internal “thinking” into the parent’s JSON feed. A dashboard or CI pipeline can now display exactly what each subagent is doing instead of a blank or “waiting” status. For operators who monitor long-running pipelines, the flag turns a black-box into a traceable step, making debugging and capacity planning far simpler.

Core stability fixes

  • Accurate status reporting – Agents now return their real state instead of fabricating results when they hit an error.
  • No auto-respawn of killed agents – Once an agent is terminated, it stays down; it no longer reappears with the original prompt.
  • Session persistence after sleep – Parallel sessions remain logged in after the host computer wakes, preventing sudden authentication failures.
  • Git worktree cleanup – Deleting jobs that live inside Git worktrees no longer triggers errors, smoothing CI/CD workflows that rely on temporary repositories.

Model handling and MCP recovery

Users of plugin MCP (multi-client protocol) servers will notice that idle sessions now reconnect cleanly instead of timing out. Model-selection bugs that produced spurious fallback notices for Vertex AI and Bedrock users have been squashed. Subagents retain the model they were assigned even after a task is resumed or a follow-up message arrives, which helps teams keep cloud-compute costs predictable.

Security hardening

  • Permission preview sanitization – Special characters that could be used to spoof prompts are now blocked, reducing the risk of tricked approvals.
  • Bash auto-mode restriction – Un-sandboxed Bash commands can no longer bypass a manual prompt, preventing silent execution of potentially dangerous shells.
  • Persistent “always allow” rules – These rules are now stored at the repository root, so they survive across different worktrees and do not disappear after a checkout.

Who benefits, who might be inconvenienced

The update wins for organizations that treat Claude Code as a server-side orchestrator—large AI-driven micro-service stacks, automated testing pipelines, and data-processing backends. Visibility into subagent actions cuts debugging time, while the tightened permission model lowers the chance of accidental privilege escalation.

What to do next

  1. Pull version 2.1.211 across all environments.
  2. Run a quick MCP idle-resume test: let a server sit idle, then restart a job and confirm reconnection.
  3. Verify that subagent model overrides survive a task resume; if they don’t, check the flag usage in your launch scripts.
  4. Observe stopped background agents to ensure they no longer emit old prompts.

Looking ahead

The current release shifts focus toward observability and safety rather than raw model upgrades.

Takeaway: Version 2.1.211 turns Claude Code from a hidden choreographer into a transparent, more secure orchestrator—exactly the upgrade production teams need to trust AI agents with critical workloads.