AI agents rebuilt a strategy game that died in 2018

Armies of Gielinor was a turn-based strategy game by Jagex, the RuneScape studio. It ran on their browser games site, FunOrb, until FunOrb shut down in 2018 and took it along. All that’s left is wiki pages and a handful of screenshots.

So I had Claude rebuild it, as a test of the gauntlet loop. That’s a technique where a fresh AI critic judges each piece blind against a reference. I wrote one prompt and went to bed. It’s playable, with 58 known defects. Next time I’d pick the models myself. The session I let choose put everything on the strongest model and used 30% of my weekly limit on that model in one evening.

Play it here. It’s a fan tribute, not affiliated with Jagex, with zero original assets. Every sprite and every sound was generated fresh.

What came out of it:

• Three separate Claude sessions ran as a relay. Prep, then the build and the judging, then shipping. They shared no memory, so each one wrote the next one’s prompt.
• The build took 8h39m overnight. One lead agent on Fable 5, Claude’s expensive top model, steered 80 others, 76 of them on Opus 5, its workhorse model. Each of those is an AI worker given one job.
• The critics flagged fewer problems every round. 380 in round one, 126 in round two and 0 in round three.
• The prep picked its own model. It ran everything on Fable 5 and took 30% of my weekly Fable limit in one evening.
• The game has its own editor for fixing what’s broken.

What the gauntlet loop is

A builder agent makes a piece. A fresh critic that never hears the builder’s side then compares it with a fixed reference, in random order, without being told which one is the original. It’s Matt Shumer’s idea, packaged by RoboNuggets as a skill (instructions Claude can load).

I wanted to see what it does when references are real but scarce, so each judged piece was tagged with how strong its reference was. I only got the per-round counts (380, 126, 0) later, from a reader I built for agent transcripts.

The title screen went through it too.

The editor

Open the editor, click anything in a running match, and you’re editing that exact thing. It’s there for the 58 logged defects and for anything the critics missed.

What it cost

Inside the build’s 8h39m, the 76 Opus 5 agents put in almost 20 hours of work. The Fable 5 lead only steered, and the whole thing fit in one night. The bill was just under 4 million output tokens (the text the agents wrote) on Opus 5 and 263k on Fable 5. I do wonder what the art would have looked like if Sonnet, the cheaper model, had drawn it. I later put Sonnet up against Opus on three long jobs.

The prep, session 1, ran entirely on Fable 5. It picked that itself, because it decided the research needed the strongest model. One evening of prep used 30% of my weekly Fable limit and 20% of my limit for all models. Would Fable steering Opus researchers have done it just as well for a fraction of that? I’d rather decide that before I send a model off on its own than find out from the meter.

The decks and the code

• Session 1, the prep and Session 2, the build and the gauntlet are narrated decks built from evidence captured while it ran. The session 2 appendix shows every image the critics judged next to what the loop made, the generated music and sounds beside what was asked for, and every agent with its model, tokens and time.
• The repo is public. It has the prompts (including the two the sessions wrote for each other), the specs, the defect lists and the retrospective: github.com/aaronorelup/armies-of-gielinor. It isn’t an ongoing project, so take whatever you want from it.