Claude Code v2.1.172: AI agents can now create their own sub-agents, 5 levels deep

Claude Code v2.1.172: AI agents can now create their own sub-agents, 5 levels deep

The headline change in this release is that sub-agents — the helper AIs Claude Code spawns during complex tasks — can now create their own sub-agents, up to five layers deep. This means Claude can break a large job into smaller pieces and handle them in parallel without you micromanaging each step. A wave of stability fixes also ships alongside this feature.

Until now, Claude Code could spin up a helper AI to tackle part of a task, but that helper couldn't delegate further. With this update, the chain can go five levels deep: an agent spawns a sub-agent, which spawns its own sub-agent, and so on. In practice, a single instruction like 'refactor this entire codebase' could trigger Claude to divide the work automatically and run many sub-tasks at the same time — a meaningful leap for solo developers building complex automation.

The bug-fix list is equally practical. A nasty issue where sessions using the 1-million-token context window would freeze permanently after running out of credits is now resolved — the session compacts itself back to a normal size instead of getting stuck. Repeated error messages about unprocessable images are gone, background agents no longer accidentally read settings from a different project folder, and Amazon Bedrock users no longer need to manually set the AWS_REGION environment variable if their local AWS config file already has it.

Key points

  • Sub-agents can now spawn their own sub-agents up to 5 levels deep — enables truly autonomous, multi-step automation
  • Sessions that hit the 1M-token limit no longer freeze permanently; they auto-compact instead
  • Amazon Bedrock now auto-reads the AWS region from your local ~/.aws config file — no extra env variable needed
  • A search bar was added to the /plugin marketplace screen, making it easier to find plugins
  • Fixed a security bug where background agents could read another project's settings by mistake

Quick term guide

sub-agent
A smaller AI agent that handles one specific part of a larger task, running separately from the main agent.
codebase
The full set of files and code that make an app or product work.
developers
Developers are people who build software, apps, or websites.
automation
A way to make repeated work happen without doing every step by hand.
context window
The amount of text an AI tool can remember and use in one chat.
Amazon Bedrock
Amazon Bedrock is an Amazon service for using different AI models through one platform.
config file
A file that tells a tool how to behave.
autonomous
The ability of an AI to complete tasks or make decisions without constant human guidance.

Sources covering this story (2)

Read original