Claude Code release notes: 2.1.217–2.1.218 bounds the agent queue

It is easy to start more agents than anyone can follow up. Claude Code 2.1.217 now caps concurrent subagents, blocks nested spawning by default, and lets a cost limit stop background work. Version 2.1.218 moves /code-review into its own background subagent. Together, the changes make the agent queue easier to manage without filling the main conversation with review work.

Source: Claude Code changelog for 2.1.217 and 2.1.218.

Claude Code 2.1.217 caps concurrent subagents

A subagent is a specialized Claude assistant that works in its own context within the same Claude Code session. It can search a codebase, analyze tests, or review a change, then return a summary to the main conversation.

Version 2.1.217 puts three clear limits around that delegation:

• Up to 20 subagents can run at once by default. The limit can be changed with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS.
• Subagents no longer spawn their own subagents by default. Teams that need deeper delegation can set an allowed depth with CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH.
• --max-budget-usd now stops background subagents too. Once the cap is reached, new spawns are denied and running background agents are halted.

Twenty is a technical default, not a target to fill. For most teams, the better question is how many results can be reviewed, tied to the right issue, and approved during the same working day.

Sources: Claude Code 2.1.217 on GitHub. Anthropic's subagent guide.

Claude Code 2.1.218 moves code review into the background

/code-review now runs as a background subagent. The review gets a separate context, the main conversation no longer carries the full review process, and stacked slash commands keep the intended change as their review target.

This does not change who starts the check. Since 2.1.215, the user has needed to invoke /code-review manually. Version 2.1.218 changes what happens next: the work runs alongside the conversation. The release also fixes descriptive arguments for /ultrareview and makes /code-review ultra launch the cloud review in non-interactive runs.

Source: Claude Code 2.1.218 on GitHub.

Diagnose connector failures without guessing

The same release makes MCP failures more visible. MCP, the Model Context Protocol, is the standard that connects Claude Code to external tools and data sources. claude mcp list and /mcp now show the HTTP status and error text when a server cannot connect. Claude Code also warns about hidden whitespace in MCP configuration values.

That is immediately useful in a real integration. A team can distinguish an authentication failure from a bad address or malformed configuration before the agent continues. Use scoped keys from environment variables or a secret manager, approval gates for writes, and a run log that ties the result to the right agent job.

Sources: Claude Code 2.1.218 release notes. Anthropic's MCP documentation.

Try this prompt this week

Human step: Update through your approved distribution path, run claude --version, and confirm 2.1.218. Open the correct repository. Check the organization's subagent-cap and budget values before changing them. Run the prompt below before invoking /code-review manually.

Source: Anthropic's changelog with version check and release notes.

Read CLAUDE.md, the agent definitions, the current diff, and saved workflow settings.
Map which subagents this task may start and where nested spawning could occur.
Calculate peak concurrency from the actual plan and flag what you cannot verify.
Propose a bounded queue and the point where /code-review should start.
Do not change files or settings.
Return a short queue receipt with owner, cap, budget, review target, blockers, and next step.

Good output should:

• Use the real agent definitions and current diff.
• Separate the technical ceiling from the team's chosen working limit.
• Name who receives each result and what still needs a human decision.

When several Claude agents work on the same codebase, the queue needs as much thought as the prompt. This is a practical Tool Forge task: set capacity, budget, and the review point before connecting more jobs.

The npm registry confirms that 2.1.218 is the current published package version.

Source: Claude Code 2.1.218 in the npm registry.