First impressions of GitHub Copilot's Agent mode

tl;dr it works. You should try it.

I finally got around to trying the new Agent functionality in GitHub Copilot.

You gotta really know when to ask an LLM for help, though. (I wrote last year how I decide whether to reach for an AI assistant, if you're interested.)

My experience and recent research both indicate that AI works best for creating new stuff from scratch, especially when that code is perfunctory and typical and conventional. That's the reason there's such a divide between people who've had terrific experiences with "vibe coding" and others—like me—for whom all the hours fucking with AI to make it do my job for me have mostly been a waste of my time. Every other time I've tried to use an agent mode in Cursor or Claude Code, I've asked it how to do things that most "staff" or "principal" or "distinguished" engineer have never tried and would have no idea how to do, so naturally they never went well.

But today I'm finally starting work on a new gem, and I had an annoying task that wasn't a legacy rescue, or exploiting a brand new language feature, or doing something totally unlike anything anyone has ever done before. So I fired up the agent mode using Anthropic's 3.7-sonnet model and let it go to work.

Specifically, I asked it to stub out a Rails engine to use as a starting point for moving over existing login functionality into a gem. It did basically everything in the same order and way that I would. It commented out additional hooks I'll probably need, which is nice.

It also gave me instructions for wiring up the new engine from the project I'm extracting the behavior from, which is a nice bonus.

Did it work? Yup:

Cool.