Claude Code release notes: 2.1.172–2.1.173 nested subagents

Claude Code had a practical release night: 2.1.172 lets subagents spawn their own subagents up to five levels deep, and 2.1.173 cleans up two details that could otherwise trip Fable 5 and Windows sandbox workflows. That sounds technical. For small teams, the better question is simpler: if Claude gets more room to delegate, do you also get a clearer work receipt?

Source: Claude Code changelog, GitHub v2.1.172, GitHub v2.1.173

Claude Code release notes: what changed

Claude Code 2.1.172 makes the subagent tree deeper. A subagent is a specialized Claude Code assistant that handles a focused side task in its own context and returns the result to the main conversation. That delegation can now continue across several layers, up to five.

The same release adds a search bar in /plugin, more model detail in the OpenTelemetry metric claude_code.lines_of_code.count, better AWS region handling for Amazon Bedrock, and fixes for background agents, model selection, wildcard rules and long conversations.

2.1.173 is smaller, but worth noting if you are testing Fable 5 or using Windows sandboxes: Fable 5 model names with a [1m] suffix are now normalized away because Fable 5 already includes 1M context by default, and a false Windows sandbox dependency warning has been fixed.

Source: Claude Code changelog, GitHub v2.1.173, npm @anthropic-ai/claude-code

Why nested subagents matter for Nordic teams

More delegation is not automatically better. It becomes better when every layer has a clear job: what the agent may read, which tools it may use, what summary it must return, and who approves the next step.

That is where release notes become practical. A development team can have a main agent ask one subagent to inspect test suites, while that subagent splits the work into narrower checks. A business team can use the same pattern for document flows: one agent gathers sources, another finds decision points, a third proposes the next action. The point is not to make the chain long. The point is to make the chain reviewable.

MCP, the Model Context Protocol, is the standard that connects AI to external tools and data sources. When Claude Code works through MCP servers, plugins or internal files, each agent branch should have scoped permissions, redacted logs, approval gates and keys stored in environment variables or a secret manager, not in chat.

Source: Create custom subagents, Run agents in parallel

Try this prompt this week

Human step: Check the version through your normal routine, for example claude --version as documented by Claude Code. Put the 2.1.172-2.1.173 changelog, relevant subagent settings and the current task next to Claude Code before asking the agent to plan.

Read our Claude Code configuration, subagent settings and latest run log for [workflow].
Propose a small subagent chain for [task].
State max depth, tools per agent, files each agent may read, and the summary the main agent should get.
Flag which steps need human approval.
Do not change files until I approve the plan.

Good output should show:

• how many agent levels are needed, and why it should not be more
• which tools and data sources each agent may use
• where logs, sources, and decisions will be saved after the run
• which changes need approval before they run

Hammer angle: do not build an agent tree without a receipt

This fits Tool Forge: connect Claude to real files, tools and systems, but make permissions and review easy to understand. Start with one workflow that already takes time today. Give the agent read access first, use scoped API keys, redact sensitive excerpts in logs and route changes through an approval gate before anything writes back.

Then nested subagents are not a black box. They become a way to split the work without losing control.