Better multi agent orchestration

Ideas Feature Requests
,
1

Feature request for product/service

Cursor IDE

Describe the request

With Codex 5.3 i found myself leaning more into orchestrating agents rather than actively being in the loop, i found Codex 5.3 strong and long running tasks, using tools and even take notes for itself as “memory”.

I currently have a qa-manager command that have a bunch of instructions on how to use Playwriter and the qa-eng subagents, also instructions on how to self-learn, take notes, ask feedback from subagents etc.

However i’m hitting significant limitations in Cursor.

  1. Subagents cannot spawn other subagents
    This is the biggest issue, i cannot have an implementation session and then spawn a qa-manager that uses qa-eng to test and take note on how to test the app correctly, i’m forced to work in separate chats making things extra clunky.

  2. Subagents cannot share the chat with the parent agent
    This can be good for “explorer” subagents like the one built-in but in case of complex subagents like “qa-eng” this is also a problem, using the browser and webapps is extremely tough for AI and they end-up in issues, loop, they struggle to understand how the app even work, there must be a way for subagents to provide extra info to the parent or the parent agent should be able to inspect the transcript to find issues and understand how to improve its prompting skills

  3. agents and subagent are isolated and unaware of all the system around them
    It’s difficult to make agents cooperate properly as they do not understand that their chat is lost after the run is finished or that subagents cannot access the same information that the invoking agent already have in its context, they do assumption, they do not share and you end-up writing way too many skill/commands/subagents files to make all this machinery work, everything feels clunky and brittle.

  4. subagents cannot ask for help or for more info
    If the agent and the subagent could communicate bi-directionally they could share informations and they would be able to cooperate better, yes you can kind of do it by using hooks and writing stuff in files and asking agents to read those files via skills but everything is clunky and it’s hard to even understand if all these skills have a positive or a negative effect on the system.

I would like to have better tool to create a proper team of agents that can communicate and cooperate, this is a necessary step before making “long running agents” a real thing, 274K tokens windows can fill up in 15 minutes and at this stage the current system works very well if you are in the loop, but completely fall aparts if you want a team of agents splitting work for “research, implementation, qa” to optimize their token usage and avoid context rot.

Operating System (if it applies)

Windows 10/11

1 Like