Paul Coroneos

I've been using the new Copilot agent mode to try to migrate my blog from NextJS App Router to Astro. I write all my blogposts in MDX with styling predefined via TailwindCSS. I then have a little legacy Webpack code that inserts getStaticProps calls where needed to fetch the blog data and then I decorate the post with a wrapper so that it looks consistent (probably should have moved a lot of this to a layout.tsx file, but you know how tech debt works).

After two 1.5 hours sessions I am pleasantly surprised to see actually most of the blog is moved over! Styling is not correct though and I will probably need another session to get this fixed. Here is my feelings so far:

Pros:

Claude 3.7 did an excellent job of translating my next.config.js to the astro.config.js. That being said I had to explain certain things such as my redirects object for it to move it to the appropriate place.
Being able to pause the agent "mid flight" was INCREDIBLY useful to course correct. I don't think the models are at a point yet where you can just let them iterate freely for hours, even if you give it a structured prompt
I was able to continue working in the background on learning system design while the code worked. I will read every line it changed before cutting traffic over, but effectively I was able to somewhat parallelize working on two things which I find quite challenging on a day to day basis.
After 3 hours of work I actually had an almost complete working clone of my blog in Astro with significantly less tech debt!
Even though copilot in agent mode went off the rails a bit on the styling... I actually like some of the spacing changes it suggested. I may accept some of these before releasing.

Cons:

Sometimes it's really, really wrong. For example I asked it to use my css files and tailwind directly and re-apply it to the new blog. It kind of went off the rails and started re-implementing what tailwind was already doing. I even fed it my existing production blog but it continued to just add more CSS rather than try to refine. This is a common complaint i've seen with current AI models.
The tools really need specific prompting and aren't nearly as context aware as expected. For example I use pnpm as my package manager but the tools assumed I wanted to use npm.
You eventually will run out of credits. After about 1.5 hours I hit a rate limit and had to start the next day. I am hopeful this reduces over time.
You really have to be specific in your prompting or it will take a lot of liberties. Those liberties will often not match what you expect

Overall I am SUPER impressed with how far along these tools have come. But it has reinforced my conclusion that in the current incarnation these tools are best served as a multiplying layer rather than a full developer replacement. I think for devs who are familiar with system design and general coding practices there is truly something here.

More to come!