Sub agents - Anthropic - I love this CC feature, would love it if you guys could bring it into Cursor - I want the ability for my main agent to spawn background agents with custom prompts + contexts + tools.
Eg: while working on a database task, lets say my agent encounters a frontend bug. Instead of fixing it itself - it should hand it off to a background agent with the relevant context and user defined prompts and MCP servers and just read the final output so we don’t bloat the main agent’s context.
@deanrie could you please confirm if is it planned for the next releases? The Codex that was recently released is already working on it and I believe it will be released in a month. Maybe their PR can help with something? PR: 3655
This would be awesome! In the meantime I have found a workaround with Cursor CLI.
You can use Cursor CLI headless mode to make prompts scriptable like: `cursor-agent -p “your prompt”`. You can combine that with “&” which makes a background process so you can do something like this in a bash script:
# This first agent is your researcher. It puts all the information needed for the
# sub-agents into a md file. We don’t use “&” so it is blocking.
# Depending on the complexity of the task you might need multiple agents here.
# Don't add the "&" at the end of this command, we need the research to complete first.
cursor-agent -p --force "Compile a list of movies released to Netflix in the last 10 years. Include the title, summary, director, actors, year it was released, and rating. Write this to the movies.md file.
# Background agent processes that perform specific tasks based off of the
# movies.md file. Treat that output as your context/memory for subagents
# Add the "&" at the end of each cursor-agent call
cursor-agent -p --force "Using the information in movies.md give me a list of the top 10 rated movies that were released on Netflix. Write this to topMovies.md" &
cursor-agent -p --force "Using the information in movies.md compile a list of directors that have worked on Netflix movies. Write this to netflixMovieDirectors.md" &
Need to have different agents profiles so for example I can have a Design agent where whenever frontend is modified, this Design agent uses Playwright MCP to review.
I can have the main agent always call the design agent whenever there’s a frontend change.
This design agent has rules and all the contexts i gave it for what’s good design for this project, and it has access to the rifht tools to do its function (such as access to browser or Playwright MCP)
Also bumping this. I was recently auto-banned from Claude, so I am looking at other options while I wait for their team to review. I’m very impressed with some of the things Cursor does.
Subagents and skills in CC are a game-changer for managing context. It was possible to get a single instance of Claude to maintain relevant context for much larger tasks by using skills to teach Claude how to delegate context-heavy work. I could also customize subagents to my codebase, which greatly increased the quality of work.
My claude.md file would be nothing but instructions on how to delegate, and then when I started a CC web instance, it would just start delegating to the subagents and managing their responses to complete the task.
Have you tried asking it to build it out for you? I’ve had it build an mcp tools registry that creates a script that creates the tool it needs to limit having too many tools. I also told it to make a skills like claude code and it did that too. It might not work exactly the same as cc, but it think it’s doing what I ask?
I basically get around this by using commands and then in the commands I tell the agent to prompt the user to run another command. It reliably does this unless the chat gets too long and then it starts to forget the initial command instructions.
The other way I get around it is in my planning command I have it split the plan mds into chunks and then I can run multiple agents at the same time to really speed up the implementation time.