Before you get mad, I've been that obsequious contract worker 🏴‍☠️

Overall, agentic "programming" feels very much like leading a team of obsequious contract workers who live in a different country, and who all have the entitlement of knowing they're getting paid no matter how long the work takes or how bad of a job they do. Not sure we needed AI for this 😅

Unquestionably, there's slowdown as things get more complex. I'm going to have to give it another run through to see how things go when starting the project greenfield. Still comparing the level of interaction and coaching needed between this and the other tools is pretty stark.

I used Warp demo and got about 65% done. Cursor got it finished but it looks like microwaved spaghetti and I wouldn't want to maintain it. Junie broke my heart. Thanks, Google, finally there's a Firebase tool I might stick with.

Basically, README for usage, .agents/spec.md for agent-facing spec, .agents/plan.md for current task specification, be *very* careful about discrete testable units when breaking work up before even turning on agent mode, and YOLO. Basically code review from that point.

I'm using poor man's spec driven development and will post examples once this is all over. It's been fun.

Firebase Studio/Gemini took a largely broken project Junie worked on for two days, roughly $25 worth of credits and has already resolved all the hallucinations into functional code.

Amazingly, it did so with almost negligible prompt changes (s/Junie/AI Agent/g, update filenames and links).

Yeah, there's significantly less friction involved here, and I never even gave it my PAYG keys. If this project was something sensitive I wouldn't be going this route, but I am just comparing tools.

I loved RubyMine when I was working in Ruby, using jetbrains was a no-brainer when I switched to python & Go, but I just can't justify sticking with jetbrains if agentic dev is this much worse than competing IDEs like warp/cursor

even if I really *really* don't like living in vscode or its variants