I've discovered a new hobby. It involves creating non-trivial libraries from scratch. (Not UI slop!) Except I don't write any code in editors. An agent does all of it.

It goes like this:

Set up an infinite loop so the coding agent that...

1. has a clear guard for correctness (unit tests)
2. has an objectively measurable, pass/fail outcome (performance benchmarks)
3. has tools for styling (dead-code tool, formatter)
4. leaves artifacts for future instances (commit changes on a branch, writes failure_analysis_SHA.md)

... and watch number go up.

You'll need a robust prompt, then just feed it to a single agent, over and over. No need to convince the agent instance it needs to repeat (doesn't hurt to try, but they usually give up after a while).

It's more fun if you ask for a success artifact (benchmark_success_SHA.tsv) and have a script that monitors for new ones. I run this script with @steipete's VibeTunnel so I can watch the number go up on my phone.

It may take some setup to get going. For example, at the beginning you want to import a large number of unit tests. You'll want to source some input for benchmarking. You'll want some infrastructure (SwiftPM project, Scripts directory, git).

But quickly, it becomes an idle game. At first, you'll watch the number of unit tests go from 100% fail to 100% pass. Then, the performance benchmarks. Bonus points if you benchmark against existing libraries, and beat them in performance. This has happened on every project I've done.

Codex btw.

I should clarify: this is not serious software engineering. The goal is having fun. Although the end result usually is really useful, and because of robust unit testing + benchmarking, it might be good enough to use in production.

We'll find out.

My favorite idle game in 2026 so far.