Lane/Orchestration Mode - Auto handoff and agent multithreading
What variant of Codex are you using?
App (MacOS)
What feature would you like to see?
I currently split my work into multiple lanes, each one consisting of a separate agent (planner, builder, tester, integrator). For the build agent, I create a new one with each larger feature set and I always need to begin the agent with rules that should persist across all build agents that I create under the project. Would really be helpful if I could store the instructions once and select it when creating a new thread/agent in that project.
EDIT: I think this might be better suited as something like a "lane/orchestration" mode, where you can create a single agent that manages agents under it for the same task type (ie. planning, QA, integration, building). Id imagine you could use it in either auto or manual. In auto, it could split a single task efficiently across multiple agents and then provide you with a single output. In manual, you could start a new agent and it would automatically add context/instructions for that lane for you. Meaning as you continue to use the lane, the context it provides agents would evolve automatically so that you get better quality results.
This would essentially give you "multithreading" for agents working on the same task, which has potential for more efficient processing, faster execution, and more thorough implementations.
Additional information
Would also be great if there were tags for agents which displayed next to the title. This way I can have a unique title instead of "Whatever Feature Builder" and just have a "Builder" tag to easily identify this.
6 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Not the same, this is a feature within the feature specified above.
This feels like more than a convenience feature. Once people naturally split work into planner / builder / tester / integrator lanes, repeated rule injection is usually a sign that the workflow wants to be project-scoped, not thread-scoped.
In many cases yes, however on projects with a long lifecycle its not feasible to use a single agent for each lane for the whole duration. Now that I think about it some more, what might make more sense is to have an "lane/orchestration" mode, where it essentially lets you create an agent that orchestrates other agents for the purpose as needed. Meaning, you wouldn't have to manually start a new thread, provide instructions/context, etc. Could even use this type of feature for "multithreading", for example the orchestration agent splits the task up efficiently across 2-3 agents and then it puts it all together. It would also be great if you could manually start an agent under the lane, say if you were doing QA on the same project but wanted to use separate agents for separate features.
I updated the title and description with a more evolved version of the concept
This feature request hasn't received enough upvotes, so I'm closing.