I feel like we've all been building a lot of things with AI that we haven't really understood deeply.
It used to be that you could only build things you'd figured out how to build. Figuring it out was a big part of the work. Now you can get something working and keep going without ever really spending time with it.
I've been doing this a lot. I've been a dev a long time, I've gone all in on coding agents, and there's so much I want to build. I don't want to stop every time the agent does something I couldn't have done myself. But I also want to come out of this a better developer.
That's how I ended up building LearningHarness. It's a learning agent alongside your coding agent. It follows what you're working on and finds things worth understanding. You can get into something when you have a moment, work through it, try it yourself. The main agent keeps going.
Hey Product Hunt, I've been a dev a long time and I've gone all in on AI coding. I've built a lot of new primitives with agents that I don't feel I've even internalized yet. They're working, and I'm already building on top of them, but my understanding hasn't caught up.
Figuring things out used to be part of building them. Now the agent gets me past the difficult bit and I keep going. Coming back to learn becomes another thing I have to remember to do.
So I built LearningHarness. It follows your coding agent's work through a read-only connection and offers things worth understanding while the agent keeps building. You can explore an explanation, work through the reasoning, or try a changed case yourself.
I don't want to slow the agent down or limit it to what I already know. I want to put some of the same token allowance toward getting better myself.
It runs locally, uses your existing model connection and allowance, and needs no extra API key. Your learning history stays on your machine. There's no LearningHarness backend receiving your code, prompts or learning responses. Usage metadata sharing with PostHog is on by default and can be turned off in settings.
It's early: macOS, Codex desktop early access, and an experimental Claude Code version.
If you've felt this gap, I'd love to know whether the learning it offers helps you understand the work more deeply, and where it gets in your way.