What are your thoughts on multi-agent coding systems. I feel like I'd get better performance if I had separate debugging agents from deployment agents but I'm not sure how to do that all in a single rules file
Yeah excellent point Ken. In an ideal world, a single agent develops and tests iteratively, but as of today I'd say there are still some advantages to having a specialized testing agent (due to today's models' instruction following limitations).
However, these advantages are pretty slight IMO and I've been having the same agent/rules file used for both, with Sonnet 4. It just makes the most sense to have the agent develop and test as it goes (like the best human devs). AFTER that though, I think there will be long-term value in having a buyoff agent that performs a test from scratch, like a human QA buyoff. So (testing+dev), and (final test). I actually believe this will come from the (future version of the) IDE, not the CLI agent. Got a project in the works to help with that.
And don't get me started on agent swarms 😵💫 a swarm is the opposite of ideal.
What are your thoughts on multi-agent coding systems. I feel like I'd get better performance if I had separate debugging agents from deployment agents but I'm not sure how to do that all in a single rules file
Yeah excellent point Ken. In an ideal world, a single agent develops and tests iteratively, but as of today I'd say there are still some advantages to having a specialized testing agent (due to today's models' instruction following limitations).
However, these advantages are pretty slight IMO and I've been having the same agent/rules file used for both, with Sonnet 4. It just makes the most sense to have the agent develop and test as it goes (like the best human devs). AFTER that though, I think there will be long-term value in having a buyoff agent that performs a test from scratch, like a human QA buyoff. So (testing+dev), and (final test). I actually believe this will come from the (future version of the) IDE, not the CLI agent. Got a project in the works to help with that.
And don't get me started on agent swarms 😵💫 a swarm is the opposite of ideal.