Boris Cherny just shipped an experimental tweak to Anthropic’s terminal-based coding agent that lets a main agent spin up subagents which can themselves spin up more subagents, creating a controlled hierarchy capped at five levels deep to keep long-running or multi-part tasks from flooding the primary context window.
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
Each subagent runs in its own isolated window and returns only a summary, so developers can now break sprawling codebases or debugging sessions into layered delegations that previously had to stay flat.
The hard limit is presented as a starting point rather than a final design, with Cherny explicitly asking users for feedback on how the nesting behaves in real workflows and whether the ceiling feels too tight or just right.
Positive users praise Claude Code's nested subagent support for better context management and specialized workflows, while negative users point to bugs, rate limits, and token costs.
“agents kicking off agents as a way to better manage context”
“nested” language models
“depth=N”
neat :-)
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
paging @kieranklaassen
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
One idea I’m experimenting with: add "fork:true" to a skill’s frontmatter to have it run in its own context window, then in the skill tell it to use agents to keep context isolated for each step also.
Adding this to the built in /code-review skill to improve performance even more
@bcherny Testing this out as soon as I have access to it. Subagents are really underrated especially for effective context management.
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
oh boy
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
@lateinteraction It's a rule that Anthropic has to rename everything that they introduce 😀
“agents kicking off agents as a way to better manage context”
“nested” language models
“depth=N”
neat :-)
@omerfrkdemiral Say “use a workflow”. Just “workflow” had too many false positives
@bcherny i love the names my codex gives it's subagents 🤣
@bcherny Previously, when I wrote workflows, they were colorful; could you bring that back?
From my experience this doesn't work out of the box. CC struggles on runaway context in a single instance and often adding more nested subagents makes it worse up to a certain point. It also really depends on the complexity of the problem, so you have to scope its use well or it can be less context efficient and actually do worse in a lot of cases. There is some coordination magic that can get it to work in the right circumstances, but that's for the CC team to figure out ;)
@bcherny like the context management angle
more agents kicking off agents = more review surface area
@plannotator feels timely
@bcherny Will you support nested dynamic workflows, too? Would love to explore that. As far as can tell, those were hard limits with tools unavailable to subagents.
@bcherny just call it RLM already
@bcherny Yea I've been /batch and loop and I used to use it a lot for monitoring but I switched to gpt max plan for that as -p is kind of less economical these days because of those rules .. but apart from that I love sub agents so cool they are back and deeper than ever :D
@bcherny Now tracking what agents did and improving will get terrible.
But Superdense can help https://github.com/Nimrobo/superdense
@bcherny Can you share an example of how you personally use it?
@bcherny Does this include forked?
@bcherny Can we watch what each subagent is doing? Monitoring subagents should be fun!
@SeekerClaw Yep!
@bcherny Sounds smart
