AI Driven Code Reviews - Sub agents
Cursor recently made it easier to invoke sub agents. I tried it out for the agentic code reviewer I blogged about recently.
- Pros: it would help manage the context & tokens (in theory) even better.
- Cons: I’d be tying the implementation more to the IDE than before. I’m trying to keep all my AI prompts IDE-independent.
The approach I tried:
Actual outcome
- I didn’t see a noticable difference either in context usage or in speed. If anything it was slower.
- In order for the orchestration to work, it had to create several more files than before & this actually made the process slower.
- Several explicit instructions were ignored in the interactions between agents.
Conclusion
It seems like sub-agents are not a good fit for my needs here. Perhaps skills might meet my needs better?