Thank you this is the clearest explanation I’ve seen of Claude code sub agents.
Regarding keeping the orchestrator / main thread pure so it always has a clear idea of your plan: have you experimented with documenting the requirements and specifications in a markdown file and having the orchestrator reference and update that?
This seems like it would harden the clarity of the plan even more and prevent accidental drift.
I’ve played around with Kiro from AWS, which has this spec driven development approach baked into the workflow, and it seemed promising as an approach for creating continuity across limited context windows.
Thank you for this insightful article! The idea of parallel spawning of sub-agents in Claude Code is completely new to me. It’s impressive how this approach can transform AI from a simple assistant into a real orchestrator for complex development workflows.
one question. Do you think I can result in the same 'separation of responsibilities' simply by having different claude code running in different sessions/terminals. i think it avoid context rot.
i understand sometimes a sub-agent, where i only give specific permissions and context can make sense, but I feel in most cases it's overkill.
also if we take this simpler approach, there isn't a need for an orchestration agent
Regarding the token costs, Haiku 4.5 is three times cheaper than Sonnet and is now capable of doing these smaller custom tasks in a way that Haiku never has been able to do in the past. Meaning: spawning armies of sub-agents is cheaper than it ever has been before!
This is brilliant! Parallel orchestration feels like the missing layer in AI-assisted dev. It’s not just faster, it’s structured intelligence in action.
For more AI trends and practical insights, check out my Substack, where I break down the latest in AI.
Thank you so much Jeff! It was a pleasure working with you.
Exceptionally pragmatic and refreshingly thorough, mate! Also, really appreciate that this piece reads like person wrote this.
Thank you this is the clearest explanation I’ve seen of Claude code sub agents.
Regarding keeping the orchestrator / main thread pure so it always has a clear idea of your plan: have you experimented with documenting the requirements and specifications in a markdown file and having the orchestrator reference and update that?
This seems like it would harden the clarity of the plan even more and prevent accidental drift.
I’ve played around with Kiro from AWS, which has this spec driven development approach baked into the workflow, and it seemed promising as an approach for creating continuity across limited context windows.
Thank you! I do this split between two files - Claude.md for overall project breakdown and Plan.md from planning mode to track progress.
Yes I also use plan md files in my more advanced workflows for complex tasks.
It depends on the complexity of the task at hand, I use a routing mechanism to respond with the appropriate workflow for the task at hand
Thank you for this insightful article! The idea of parallel spawning of sub-agents in Claude Code is completely new to me. It’s impressive how this approach can transform AI from a simple assistant into a real orchestrator for complex development workflows.
Great article.
one question. Do you think I can result in the same 'separation of responsibilities' simply by having different claude code running in different sessions/terminals. i think it avoid context rot.
i understand sometimes a sub-agent, where i only give specific permissions and context can make sense, but I feel in most cases it's overkill.
also if we take this simpler approach, there isn't a need for an orchestration agent
Regarding the token costs, Haiku 4.5 is three times cheaper than Sonnet and is now capable of doing these smaller custom tasks in a way that Haiku never has been able to do in the past. Meaning: spawning armies of sub-agents is cheaper than it ever has been before!
Heck yeah, Haiku is great for information gathering and Sonnet for planning with said information
Awesome thank you so much for the detailed guide, especially with the slash commands! I'll totally be trying that out ☺️
Thank you so much Jeff!
This is brilliant! Parallel orchestration feels like the missing layer in AI-assisted dev. It’s not just faster, it’s structured intelligence in action.
For more AI trends and practical insights, check out my Substack, where I break down the latest in AI.