1. mainly for discovery and research. I clone all the repos I need and then I talk with the agent: "I need to work on service A which has dependencies on service B, C, D and E. How can I accomplish X? Oh ok it seems, B can offer that but only when Z happens, after that I should rely on C. Cool. And now we probably need to extend E. Umm, ok, but I think E's team won't like that. Hold on, I'll reach out to them. Give me a succinct paragraph explaining what we need from them. Alright I'm back, so we have green light but only if we implement it like H. So, give me a plan"
2. Implement the plan. So the agent spits out code that I need to babysit (e.g., perhaps the class is not in the best package, or perhaps we are adding too many e2e tests when unit tests could be enough, etc.)
3. Answer product questions like: "Why is button K disabled in this screenshot? What are the potential causes? Afaik, I have the permissions A, B and C, and I have enough Z, so why is still disabled?", "How does the flow for endpoint 'api/a/b/c/ work? Where are the entry points? Dependant services? Tables touched?"
I could live without the agent generating the code, but not anymore without their help for discovery and research. I dislike using agents for coding (I feel the same as you), but not for research (I actually use them as a sort of improved "google for everything")