Skip to content

Product hardening is finding the seams

It has been an absolute marathon the last few days.

I’ve been finishing the coding agent and doing the product hardening work, which really means trying to keep pace with the rest of the industry.

About 150 tickets’ worth of things I knew I needed to do, but had put off. All the unglamorous, necessary work that takes something from a useful demo to a plausible product. The gulf there is bigger than most people realize.

A lot of new features went in. A lot got tested. I recertified the models I’m working with half a dozen times, because anything I do to the core loop means a retest. That’s probabilistic systems. No way around it.

But after all the bug fixes and feature work, testing got much faster. What used to take about two hours per model is now about fifteen minutes for the entire workflow.

And yes, fifteen minutes still sounds heavy. But it really isn’t when you take into account how much this system is doing in the background.

Less task switching, more autonomy

Operationally, the goal with this project was to stop the task-switching headaches I’ve had pretty constantly since I started working with AI systems.

A lot of coding agents today want you living in six windows at once, constantly monitoring them, constantly deciding, constantly reconstructing what’s happening. It’s fun. It’s also cognitively expensive.

So I wanted something more autonomous. Something I could work with more like an employee than a task runner that needs constant intervention.

The decisions still matter. They’re just top-loaded into planning, instead of being spread across a thousand cyclomatic micro-decisions that, let’s be honest, most of us weren’t really paying that much attention to anyway.

It’s less “set it and forget it” and more plan → produce → review.

And it works well.

I’ve been building agent frameworks for three years now, and this is by far the strongest one I’ve come up with so far. Slick, minimal, solid TUI.

And when I’m done, it’s going to connect to Discord and Slack.

I’m tired, but I’m close.

And I’m still pretty excited about the project.