Claude Code adds nested subagents to keep context under control

Claude Code has just rolled out nested subagent support, an experiment in “agents kicking off agents” to better manage context. The first release caps nesting at depth=5, with a new “fork:true” option to run skills in separate context windows.

•
Claude
claude cover

TL;DR

  • Nested subagent support: Experimental “agents kicking off agents” for context management; capped at depth=5
  • Release timing: Feature shipping in “today’s release,” per Boris Cherny
  • "fork:true" frontmatter: Runs a skill in its own context window; supports step-level context isolation via agents
  • /code-review skill update: Incorporating forked-agent approach to improve performance
  • Practical notes: Subagents can use Chrome; forked subagents supported; terminal thread viewing via “Hit arrow down”

Boris Cherny posted on X that Claude Code has gained "nested subagent support," which he described as an experiment in "agents kicking off agents" to manage context. Cherny said the feature is "capped at depth=5" for now and is "going out in today’s release."

In follow-up replies, Cherny pointed to one approach he is testing: adding "fork:true" to a skill’s frontmatter so it runs in its "own context window." He added that the skill can then be told to use agents to keep context isolated for each step, and that the built-in /code-review skill is being updated with that idea to improve performance.

Cherny also answered a few practical questions from commenters. He affirmed that subagents can use Chrome and that forked subagents are supported. When asked about viewing multiple agent threads in terminal, he replied, "Hit arrow down." Other commenters asked about rate limits and nested dynamic workflows, but Cherny did not offer details on those points in the thread.

Source: Boris Cherny on X

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community

Related Articles

claude cover

Anthropic launches Claude Fable 5 and Mythos 5 models

Anthropic has just rolled out Claude Fable 5 and Claude Mythos 5, pairing a general-release model with a more permissive version for trusted users. The company touts big gains in coding, vision, and research, plus new safety routing and lower API pricing.

1 shared tag
claude cover

Claude Code’s /fork now runs background agents with full context

Anthropic says Claude Code’s /fork can now spin up a background agent using the exact context from your active session, then return results to the original thread. It also clarifies /branch as the older, manual workflow—sparking questions about tokens and receipts.

1 shared tag
claude cover

Anthropic launches ant CLI to run Claude APIs in terminal

ClaudeDevs has rolled out a new ant CLI for Claude Platform, bringing Messages API and Managed Agents to the terminal. It promises OAuth login, piped YAML requests, TUI exploration, and Git-versioned agent configs synced from CI.

1 shared tag