Build a browser game with a self-improving loop of agents

Juri Kern set up a loop of agents to build Depot Script, a browser game where players write short Python-like scripts for couriers in a parcel depot, from 16 to 22 October 2026. Each round, a researcher read the last round's transcripts and git history and proposed process changes, a visionary proposed decisions, and an orchestrator wrote briefs for designer, engineer and marketing loops. Each loop paired an author agent with a read-only attacker that ran mutation tests. A four-seat panel, two Claude and two on the OpenAI Codex CLI, voted on every game change, and a controller merged only diffs whose hash matched what was voted on. Novice agents playtested builds in a real browser. Claude Opus handled judgment, Sonnet the drafting and attacking, and a local model through Ollama did free first drafts. Kern answered owner questions, approved deploys and ran anything touching production.

The loop ran 54 rounds and made 338 commits. The share of changes passing review on the first vote rose from 49% to about 90%. Two browser playtest agents finished the game. When Kern later added an outside score, fresh agents that never saw the code, none of 18 reached goal 5 in a text harness, a wall no inside metric had shown. A DeepSeek red-team lane cost 1.05 USD for zero usable verdicts and was stopped. A planned Steam release was dropped and the demo went to itch.io.

Add the outside measure at round one and hash-pin the graders, or the loop will happily climb its own scores. The rules file grew to about 330 rules in a week, so he now expires rules after 14 days unless cited. The agents re-read about 690 cached tokens per token written, which is why he kept the orchestrator on one model. The harness running the Claude roles is not named.