Buildat in 2026

I recently (2026-09) took up on building upon and fleshing out my second game/software platform Buildat. I'm developing it mostly privately, but I'm posting squashed pre-releases with source and binaries on Github. Feel free to try it out - it's quite interesting and promising I think. It's the very bold do-it-all big brother to Luanti.

I originally started Buildat in 2014, and left it as it was too much work at the wrong time. Inspired by new Luanti forks popping up this year (2026) and figuring out it might be Time(tm), I took a new look at it.

Buildat was still up in the open at Github, with no users or contributors, and decided the architecture still stands. I tried fixing the build using an AI agent and that went well. Then I laid out a plan to fix some annoyances and flesh out the rendering, and put the agent at work again, and that went well too.

Up until today as I'm writing this (2026-10-03), I've been appending and updating the plan, and letting the agent continue implementing All The Things(tm). I'm maintaining a software developemnt methodology where scripted playtesting plays a major role, which allows the AI agent to iterate on Buildat at its own breakneck pace. I've found out that as long as I poke my head in doing some playtesting once in a while and keep the plan populated and maintained, this process converges into nicely working new features.

When I notice any slop, I call stop, figure out why it's happening, and sort it out. Usually the reason is that I thought I planned something well, but actually didn't, and that causes the agent to lack a clear goal, which causes it to start laying out slop.

That does make sense, though. Set the plan up in such a way that it converges either directly based on what you wrote, or as the combined result of your goal and the reality available to the agent. Existing code, available tools, metrics, docs and whatnot - no need to mention exactly what to resort to, the agent will figure it out - but having the right hunch about what the agent's context will hold and what its biases will lead it to is the art and skill required for pulling this off.

Due to the breakneck speed Anthropic's Opus can iterate out quality code, at times keeping up with laying out the plan requires continuous effort from me for hours on end. It's worth it because it keeps the agent pointed at the goal and I keep getting exactly what I want.

I'm estimating that compared to implementing the features, doing the playtesting rounds, investigating the bugs and doing all the other required things by hand, the productivity of this method is in the order of 10x to 100x, at the cost of around 100€/month. Yes, that's ridiculous and that is why picking up a project like this after more than a decade of non-viability can happen now.

Before you ask: Yes, I wrote this myself. Forcing people to read AI prose without making it obvious at the get go is evil. Neither did I use autocomplete or spell checking.