Claude Code 2.1.224 Drops 200-Subagent Cap: Scale Your Parallel Workflows Now

Key Takeaways

• Claude Code 2.1.224 lifts the 200-subagent cap and adds cross-session SendMessage/ListAgents.
• Update now to run unlimited parallel agents and orchestrate multi-machine workflows.

What Changed

Claude Code 2.1.224 is a major release that removes a long-standing scalability bottleneck: the 200-spawned-subagent session limit. For teams running large parallel test suites or multi-agent code reviews, this was a hard wall. It's gone.

The release also introduces three headline features:

1. Cross-session messaging — SendMessage lets Claude Code sessions on any of your machines talk to each other. ListAgents discovers them. This is macOS and Linux only.
2. Self-hosted runners — claude self-hosted-runner turns your own machines or containers into execution targets for Claude Code web, mobile, and desktop sessions (Team/Enterprise plans).
3. Archive plugin source — Install plugins from a zip over HTTPS without git or npm, with optional SHA-256 pinning for supply-chain security.

Plus a batch of security fixes, including a sandbox bypass where denyRead: "~/.aws/" with a trailing slash could be silently ignored on Linux and macOS.

What It Means For You

Unlimited subagents = bigger parallel work. Previously, hitting 200 spawned subagents meant your session would stop spawning new ones, forcing you to serialize work or restart. Now you can fan out hundreds of agents for tasks like:

• Running a full test matrix across 500+ configurations
• Parallel code review of a large PR across multiple files
• Bulk refactoring with independent verification per file

Cross-session messaging enables agent swarms. With SendMessage and ListAgents, you can now coordinate multiple Claude Code sessions across machines. A common pattern: a "manager" session that dispatches work to "worker" sessions on different machines, each with its own context window and resources.

Self-hosted runners decouple compute. If you're on Team/Enterprise, you can now run web/mobile/desktop sessions on your own hardware — useful for air-gapped environments or when you want to leverage a beefy GPU box.

Try It Now

1. Update: Run claude update to get 2.1.224.
2. Test the subagent cap: Write a script that spawns 250 subagents. Previously it would fail; now it should complete.
3. Set up cross-session messaging: On two machines (macOS/Linux), run claude in two sessions. Use ListAgents to find each other, then SendMessage to pass a task.
4. Harden your sandbox: Check your ~/.claude/settings.json for deny rules with trailing slashes (e.g., denyRead: "~/.aws/"). Remove the trailing slash to ensure the rule is enforced.
5. Try archive plugins: Host a plugin zip on your internal server and install with claude plugin install <https-url> --sha256 <hash>.

Source: github.com

[Updated 12 Aug via reddit_claude]

A new security caveat surfaced for Claude Code users: all session data — including prompts, tool calls, and results — is stored as plaintext JSON locally. [per Reddit user CoderLuii] Sessions live in ~/.claude/projects/<slugified-path>/<session-id>.jsonl (Windows: C:\Users\you\.claude\projects\). One user reported a single session file reaching 35MB, with 71 files in one project. This means anything pasted into a session, such as API keys or sensitive code, is sitting unencrypted on disk. While useful for auditing actual agent actions via tool_result entries or grepping for past commands, it underscores the need to avoid pasting secrets into prompts.