Layering/Managing Agents

Discussions
1

Right now I use the agents to develop single features - which works great! I first let it make a plan, maybe make some improvements, then let it write the code and run a quality check at the end. These are atleast 3 manual messages I have to send.

I’d like to move the ladder up one step by having an agent delegate agents for me, more like a manager than a developer. So the manager would maybe have a few different agents develop a plan, choose the best one, let it develop and hand the result to a quality assurance agent and communicates with him.

Is there a way to do this? I know subagents, but they seem very narrow. For quality control they might work, but I’m unsure of development. The subagent would need to have the same instructions as the primary agent that is in place right now.

However it feels like the subagent is more of a slim version of the main agent, focussing only on specific tasks.

Going further even the managing agents could get managers which decide on feature packets. Like in real companies, the boss doesnt decide on how every feature is implemented or how the architecture is, the employees do that by working together.

Is there something like this already in cursor or would I need to develop it myself? I was thinking maybe claude managed agents coupled with the cursor cli

3

Hey, this is exactly the kind of case that custom subagents + hooks are built for. Subagents are not limited to the slim version. You can define your own in .cursor/agents/ with your own prompt, tool set, and model. So a developer subagent with full instructions, a planner, a QA reviewer, and so on, is all doable.

Quick recipe:

  1. Custom subagents: define each role as a separate file in .cursor/agents/ planner, developer, QA. The parent agent calls them via the Task tool.
  2. Hooks subagentStop: lets you auto-trigger a follow-up. For example, when the developer subagent finishes, the hook returns a followup_message that sends the result to the QA subagent. That is the manager-style communication you described.
  3. Parent agent as the manager: in its system prompt or rules, spell out the delegation strategy like for planning call X, for coding call Y, for review call Z.

Read:

Multi-level hierarchy managers of managers is not supported natively. Subagents only go one level down, and they can’t spawn their own subagents. If you need deeper orchestration, Cursor CLI in headless mode plus an external script gives you full control, but that is a custom build.

4

If you’re looking to try something quickly, here’s an open-source agent harness based on Avatar the Last Airbender: GitHub - tmcfarlane/oh-my-cursor: Like “oh-my-opencode”, but for Cursor IDE. Multi-agent orchestration, natively, using nothing but a few config files. · GitHub