> We initially built an integrator role for quality control and conflict resolution, but found it created more bottlenecks than it solved
Of course it creates bottlenecks, since code quality takes time and people don’t get it right on the first try when the changes are complex. I could also be faster if I pushed directly to prod!
Don’t get me wrong. I use these tools, and I can see the productivity gains. But I also believe the only way to achieve the results they show is to sacrifice quality, because no software engineer can review the changes at the same speed the agent generates code. They may solve that problem, or maybe the industry will change so only output and LOC matter, but until then I will keep cursing the agent until I get the result I want.
In my experience, spending 20–30 minutes writing a good spec results in code that is about 90% close to what I expected, which reduces the back-and-forth with the tool. It also helps me clarify and define with some level of precision what I actually want. During the specification phase, I can iterate until the design proposed by the tool is close to what I envision, reducing the number of surprises when the tool generates code. It’s not perfect, and there are still details the tool misses that require additional prompts, but overall I can get good results in a single session, whereas before I would exhaust the tokens and need to start a new session again.
If you don’t mind learning another language. I have found Learn Concurrent Programming with Go by James Cutajar to be a very practical book. It includes memory sharing and message passing approaches with plenty of examples. It also explains concepts like mutual exclusion, deadlock-free and starvation-free properties and others. For Java, you can try The Art of Multiprocessor Programming Second Edition. It includes examples in Java but it is more theoretical and it includes a lot of proofs, specially the first half. The second half is more approachable.
Of course it creates bottlenecks, since code quality takes time and people don’t get it right on the first try when the changes are complex. I could also be faster if I pushed directly to prod!
Don’t get me wrong. I use these tools, and I can see the productivity gains. But I also believe the only way to achieve the results they show is to sacrifice quality, because no software engineer can review the changes at the same speed the agent generates code. They may solve that problem, or maybe the industry will change so only output and LOC matter, but until then I will keep cursing the agent until I get the result I want.