Claude Code adds nested subagents with depth=5 in today’s release

• bcherny's release post says Claude Code can now nest subagents up to five levels deep, giving agents a way to kick off more agents instead of keeping every step in one thread.
• According to bcherny's SDK note, the expanded /usage breakdown is already exposed in the SDK and is headed to Claude Desktop and mobile, while bcherny's /usage reply says the terminal view already attributes tokens to skills, MCPs, and plugins.
• Official docs for custom subagents describe the same core benefit: each worker runs in its own context window and returns a summary, which is exactly the kind of context hygiene nested delegation needs.
• The new usage view is more than a cost meter. As bcherny's debug reply puts it, the expensive part is often a bad plugin, and the cost docs now break usage out by skills, subagents, plugins, and individual MCP servers.

You can see the official framing in the subagents docs, the larger orchestration model in the dynamic workflows guide, and the product breadcrumb in the Claude Code changelog. The weirdly useful bit is that bcherny's workflow trigger reply suggests a natural-language phrase, "use a workflow," is enough to invoke the workflow layer, while marckohlbrugge's 37signals skills post shows how fast the community is already turning model behavior into reusable skill packs.

Boris Cherny, Claude Code lead at Anthropic, framed the release as context management first, not raw parallelism. In his announcement, he said nested subagents are for "agents kicking off agents" so complex work can stay isolated instead of bloating the parent session.

The hard cap is five levels for now. Cherny's follow-up calls it a hard limit, while the changelog notes that sub-agents can now spawn their own sub-agents up to five levels deep.

That matches the existing subagents docs, which describe a worker pattern where each subagent gets its own context window, permissions, and system prompt, then returns only a summary upstream.

The other quiet ship is observability. In bcherny's /usage reply, the breakdown names the culprits directly: skills, MCPs, and plugins. His terminal tip adds that terminal users can reach it by scrolling down, which makes this feel less like a new dashboard and more like hidden instrumentation that finally surfaced.

The official cost docs go further than the tweet. On paid plans, /usage attributes recent usage to skills, subagents, plugins, and individual MCP servers, with 24 hour and 7 day views. The changelog also says the breakdown now shows what is driving limits usage, including per-MCP-server cost.

That changes the debugging story. According to bcherny's support reply, the thing blowing up a session is often a bad plugin, not the model itself.

Nested delegation lands inside a bigger Claude Code pattern that was already visible in the docs and tweets this week. bcherny's five tips post lists the stack in unusually concrete terms:

1. Auto mode for permissions.
2. Dynamic workflows for hundreds or thousands of agents.
3. /goal or /loop to keep runs going.
4. Cloud execution so the laptop can close.
5. End-to-end self-verification.

The dynamic workflows guide describes that orchestration layer as a JavaScript runtime Claude writes and reruns in the background, with the main session staying responsive. bcherny's self-verification post makes the taste shift explicit: longer-running agents need their own verification loops, not just bigger context windows.

One small but telling detail came in bcherny's workflow trigger reply, where he said users should say "use a workflow" because the shorter word "workflow" caused too many false positives. Claude Code is getting more voice-operated, and the command language is being tuned like a product surface.

The fastest community response is not new agents, it is packaging judgment. In marckohlbrugge's post, Marc Kohlbrugge says he ran Claude Fable 5 over 37signals' open source Rails code, extracted the patterns into reusable guides and skills, and dropped them into a public 37signals-skills repo.

That repo is basically a style transplant for Claude Code. Its README says the skills can live in ~/.claude/skills/ or a project-local .claude/skills/ folder, and the official skills docs describe the same mechanism: prompt-based tools Claude can invoke automatically when a task matches.

Nested subagents make that more interesting. Cherny's frontmatter experiment reply mentions a fork:true skill frontmatter idea so a skill can run in its own context window, which points toward skills that do not just add instructions, but also choose their own execution boundary.

One more pattern surfaced in the replies: people are treating Claude Code as a remote runtime. In levelsio's VPS workflow post, Pieter Levels says he SSHes into a VPS with Termius, installs Claude Code there, runs it with bypass permissions, and closes the laptop whenever he wants.

That lines up with bcherny's long-running tips post, which lists cloud usage as the easiest way to let Claude keep working for hours or days, and with Anthropic's own quickstart docs linked from levelsio's reply. For creative builders doing long vibe-coded runs, the story here is simple: Claude Code is drifting away from the local chatbox model and toward a delegated background workspace.