Building with agents in 2026

December 2025

Over the holiday break, I wanted to push myself to get back into coding again. With Opus 4.5, and seeing all the cool tools other designers were shipping, I was running out of excuses not to start.

First taste is free

I played with Cursor a few times in 2025. A session or two with Sonnet 3.5 in February, then a few more in May with Sonnet 4, and checking out Composer 1 in October.

But didn't really start using daily until late December. I really miss Delicious Library. It was one of the first applications that I found gorgeous, and really made me appreaciate software design as a craft.

I have a few movie backups, so I wanted to build a Delicous Library clone to help me sort through them, tag em, and add some reviews. I ended up with libcat (opens in new tab).

Writing code on Windows

My macbook is ancient (intel, 2019) so I had to use my windows 10 setup for all my dev work, which was a whole learning experience in itself. I mostly used cmd, then powershell, and in the last two weeks learned about Windows Terminal, which is a great upgrade.

Starting off with the free usage, I used up most for the hobby plan in less than an hour. Switching to composer-1 helped with costs, but I ran into quality issues.

So I moved on to Cursor Pro and went back to Opus 4.5, and used that up in a couple more hours. Then I moved to Pro+, and that ran out a day later. So decided to try out Ultra. 20x usage should be enough to finish libcat.

My prompts were pretty terrible at the start. The more I used agent chat, the better it got. I started to take screenshots, and it would understand what I was talking about. I was getting so used to Opus getting everything correct, that I started to get sloppy, give less context, then be frustrated when the results dropped.

I manage to finally learn that context is everything. Breaking large changes into smaller pieces, seperate threads, produced much better results, with less bugs.

I was using Cursor so much, I would be running out of usage by only day 15, so I started experimenting with other models, but nothing clicked as well as Opus 4.5.

The IDE to CLI Pipeline

When I hit about 80% of Cursor Ultra usage, I was getting more curious to try out claude code more seriously. I kept seeing everyone using skills that produced great results, so I gave it a try on just random ideas I've had over the years, and it one-shotted most of them.

I was impressed. Similar to the hobby plan on Cursor, I ran out of free credits within a few prompts. Moved on to Pro to get more usage, and set myself a budget not to exceed the 20$ a month plan.

But I kept running into the usage window limits. I would run a few prompts, get in the flow, moving really fast, then get the cli notification:

you've used up 95% of your usage. Wait 4h38m to reset

So I ended up getting the Claude Max plan too, and productivity increased. I started running multiple CC instances, but was still using Cursor for some things, managing git, and small ui/motion tweaks.

I shipped a ton of repositories (opens in new tab) in 30 days. A lot of slop in there, but it was such a great learning experience. I haven't felt that motivated to jump into code in years.

Last week I went to a Claude Code meetup (opens in new tab), and it was great to see the passion in the room. Everyone from students to lawyers looking to use these new tools.

TLDR of what I learned

I need to do a better job of tracking what I'm learning (Insipired once again by Brian Lovin's TIL (opens in new tab)), but here's a bunch of things I picked up over the last month:

• ffmpeg is wonderful piece of software
• convex (opens in new tab) is very cool, my favorite piece of my tech stack
• voice to text transcription models (opens in new tab) are very good, not perfect, but very close.
• manging hosting is still a pain. althought Vercel and Railway are good options, VPS is still the best way to go
• managing auth is a pain, but so far better-auth (opens in new tab) is my favorite
• electron (opens in new tab) is fast, tauri (opens in new tab) is slow
• rsbuild (opens in new tab) is very fast, vite is still excellent
• i love bun
• cli tools (opens in new tab) are fun to build, more constraints, more focus on solving real problem
• project tracking and managing tasks is still a pain. i started with a text file, then obsidian, then i shipped eperdu (opens in new tab) to help me. Notion would of been just as good tbh
• docker is great
• caddy is a much better first time experience to set up, compared to nginx

What's next?

I have downgraded both Cursor and Claude to the Pro plans. I am being more mindful of my usage, and focusing more on craft rather than minmaxing out my usage every hour of the day.

I'm very happy with gpt-5.2 and gpt-5.3 in the codex cli (unfortunately the Codex macOS app doesn't support Intel natively). Paired with the insane speed of Zed (opens in new tab), it's been a dream. I still like to do reviews in an IDE before pushing changes, and Zed has been perfect for that workflow.

Codex has 2x usage limits until April and unless a new model drops, or my opencode black (opens in new tab) subscription gets activated, it should suit me fine.

I also want to spend more time with Kimi 2.5 (opens in new tab) and other open models, I'm very optimistic that we should be able to use these tools without spending hundred of dollars a month.

I plan to write more about my experience building some of these projects (opens in new tab), but I wanted to share my progress so far.

If you're building dev tools, let's chat.