So you say your esoteric meta-harness is pretty good, huh? You got your clankers in check and they're churning out code day and night. OK, buddy, but what are you actually building? Are you making something sensible and useful? Anything with actual users? Any work on legacy code or just greenfield?
First of all, I have numbers. The numbers are collected mechanically from all finished Gantry runs on my workstation, from more than a dozen different repos. Is that a perfect testbed for what Gantry can to in the real world? No, it's not. But it's more than most meta-harness makers will give you in terms of hard numbers.
Of course, people will also want to know what kind of projects is being worked on. So let me tell you.
The clanker controller itself
The project Gantry has been doing most work on is Gantry itself. I have so far done 178 runs of Gantry to build Gantry, a total of 5,357 agent sessions launched toward this purpose. In terms of the amount of changes, Gantry has made 76% of the changes ever made to the Gantry codebase.
A tool working on building itself is an odd special case. It's kind of cool. Very meta. And it's real dogfooding. I'm proud of the numbers I just mentioned. But what does it tell you?
Well, Gantry is a Rust codebase, about 100 kloc excluding comments and tests. It does a lot of complex things. And some of the complexity is externalized. Gantry has its own configuration language for defining orchestration patterns, so most of the control flow details is implemented in TOML files, not in Rust files.
I have also done about a dozen Gantry runs to build the Gantry website, including things like the binary/source release pipeline, a fancy documentation generator, the replay system which let's you look at hundreds of actual Gantry runs and the timelapse-style screencast videos.
The procedural music generator
The second most Gantry-built project is a procedural music generator I have been working on for a couple of years, where I use tons of music theory, seed-randomization and an intricate skeumorphic web UI to let a synthetic band play complex tunes. All of it perfectly deterministic, i.e. if you enter the same set of seeds, you get the exact same song. It's one of the most complex things I have ever built and not really close to a public release. It's actually online, so if you know the URL you can go play with it right now and listen to countless little tunes, but at the moment it's just for myself and friends to try out.
In any case, I have run Gantry 49 times in that repo. Including a gigantic monster run where I let Gantry reimplement the whole thing from scratch. I spent the better part of two weeks planning that run, accumulating 28 planning files, totalling 1.1 MB of markdown. Obviously, I did that largely to see how Gantry would deal with the challenge, and I learned a lot. And the rewrite works great. It successfully unblocked the project after regular agents had created a mess.
The game of my childhood dreams
I'm also using Gantry to develop a game, a project started in part as a challenge for Gantry. It uses the Godot engine and is a mashup of a few different genres, something like a post-apocalyptic colony sim inspired by games like Rimworld and Fallout Shelter with a deep role-playing system inspired by Crusader Kings. The setting is inspired by the 1995 movie Waterworld and the game is built in a retro style, with graphics which might have run on a computer of that era. As it happens, 1995 was roughly the peak of my own time as a hard core gamer. It's meant to play as a casual idle-style game but with unusual role-playing and resource-management depth. The working title is Castaways.
I have no real background in game development. Making a game is just an old childhood dream, and I have tinkered with game projects now and then over the years. Never released anything. Will this game make it all the way to Steam, the app stores etc? I don't know. I hope so! For now, it's an experiment. I'm having a lot of fun with it, pouring all kinds of game dev ideas into it that I have thought about and experimented with over the years.
Hopefully I will do a proper write-up at some point. I'm not shy about showing off the game in its current unfinished (but very much playable state). For now, however, I will just give some stats:
Gantry has been run 44 times to build Castaways. It has launched a total of 1,840 agents, of which 473 sessions were sprint execution. It has made 299,582 line changes in 1,956 files. The edits made by Gantry-launched agents make up 96% of the total changes to code and assets. So the game is almost completely implemented by Gantry. The rest is small fixes I've made in-session using Claude Code, for things where Gantry was simply overkill. Often minor clean-up after a long Gantry run.
Web apps, because that's my thing
I started using Ruby almost 20 years ago, in the 1.8 era. I was not originally a Rails dev, but at this point I have built more Rails apps than I can count. So, naturally, much of my work with agentic coding has been building with Ruby and often with Rails. Gantry has done work on about a dozen Rails apps of varying sizes, mostly pre-existing ones where Gantry has added features, done refactorings etc.
Perhaps I can showcase one or more such sites at some point, but for various reasons it will have to wait.
Siblings in the HyperClank family
I am also using Gantry to work on some sibling projects of Gantry, other TUI-based apps for my own productivity needs, which might or might not become commercial offerings in time. One example is an app I call Skriv, which is a text editor, an Obsidian-compatible note taking system which has been my daily-driver for almost all my writing for more than a year. Skriv was fully working before Gantry was created, but I have been using Gantry to extend it.
Another example is an email management system I am working on called Brev, which I am trying to build mostly using Gantry. It's a permanent local archive for all my mail, syncing from all my mail accounts and providing a slick TUI email client with blazing fast search and 100% keyboard driven email operations.
All my TUI apps use my own Ratatui-based UI toolkit called mcurses, which provides Rust crates implementing things like a menu system, a text editor, a file picker, spreadsheet-like tables etc. I have successfully used Gantry to refactor and extend the mcurses system.
It was primarily built for something else
One of the core motivations for building my own agent orchestration system was a problem I ran into in a startup I'm working on. The project deals with bioinformatics and I built an automated research pipeline to extract structured information from thousands of papers from scientific journals. It used LLM agents but also a lot of deterministic data validation and for every agent producing data, it had multiple other agents double-checking the output.
Up until the Opus 4.6 era, it worked great, but Opus 4.7 was unable to drive the pipeline. It went 'lazy' and 'gave up' on run after run, basically it refused to do the work that Opus 4.6 and predecessors had done without issue for months. I started breaking the process up into smaller units and removing LLM agents from the loop in matters of control flow. I kind of got it working, but it felt like a Rube Goldberg machine.
I had toyed with the idea of building an agent sequencing system which would replace all my agent-launching bash scripts and when I realized that my genetics startup was an ideal consumer, it pushed me toward actually doing it. So I took a break from bioinformatics and spent the summer of 2026 building Gantry.
At the time of this writing, Gantry has not been used for anything in that other project, but I consider it the main target for Gantry in the sense that it's the most high-value thing I myself plan to use Gantry for. And that's why Gantry doesn't hardcode orchestration patterns in compiled code – because it was never meant to be used only for software development – the flexible scheme for creating custom sequencing flows was created with the research automation use case in mind.