(Continuously Updated) My Real-Time Review of Cursor Subagents

As a result, I came up with the following set:

  • Agent: GPT-5.2 XHigh
  • Subagents:
    • DB: Opus 4.5 Thinking
    • Senior: GPT-5.2 XHigh
    • Middle+: Codex 5.2 XHigh
    • Middle: Gemini 3 Flash
    • Junior: Grok Code Fast
    • Verifier: Gemini 3 Flash
  • Hidden Cursor Subagent:
    • explorer: Composer 1

I’ll tell you more in Agent Compass when I get around to updating it.

I have got a question from your practical use, do you use any subagents to run in background? Have you figured out what happens if main agents spawns a subagent that runs in background and main agent ends? will subagent finish or terminated?

Techlead.mdc:

(...)
## Limitations and Recommendations

### Limitations
- CI is single-threaded, meaning that only one subagent can be running at a time.
(...)

On one of my current projects, I had to downgrade from Python 3.13 to Python 3.12. I tasked the Subagent with checking what was broken.

The Subagent spent 33 minutes ensuring the project passed CI correctly. Meanwhile, the Agent reported: “After downgrading, nothing critical broke.” :joy:


2026-01-29_17-54-08
2026-01-29_17-53-52

:smiling_face_with_sunglasses:

  1. Using Codex CLI, Codex-5.3 XHigh created a three-phase refactoring plan (DRY, SRP, YAGNI as soft GC).
  2. I added two verifier agents to the project, based on Gemini 3 Flash and Codex-5.2 XHigh.
  3. I tasked Codex-5.2 XHigh with fully implementing the plan, verifying itself with CI and G3F after each phase, and finally verifying the work with C5.2XH.


2026-02-07_23-30-52
2026-02-07_23-36-19

My jaw literally dropped when Gemini 3 Flash launched subagents of its own model for a task with parallelization capabilities :exploding_head:

UPD: Not repeated on a similar task: he hung several languages ​​on the subagent and also called Composer, not himself.