Ascendant's Archipelago: the game

Problem

Could one person direct a game with real systemic depth, using coding agents as the implementation team, without the codebase or the canon falling apart?

Approach

• Clear roles. The repository’s instructions tell agents I’m “the creative director, not the implementation engineer.” Agents implement and decide technical matters; canon, scope, and art come to me, and canon lives in a separate worldbuilding vault the game must follow.
• Agents that check their own work. One automation entry point, playtest scenarios that drive the real game, and verify gates a merge can’t skip. My feedback arrives the same way: one key records a screenshot, the game state, and my voice, transcribed locally and triaged into tasks with an acceptance check. The first session found the whole map was mirrored, which no test could see because every test shared the same wrong frame.
• Measure the factory. Timing everything overturned my assumptions twice: start-up was slow because of per-line log flushing, not rendering, and the limit of three concurrent Unreal processes I’d asked for was costing 55 minutes a day while six showed no slowdown. A full verify fell from about 45 minutes to 13, then climbed back to a median of 44 as the suite grew to 49 scenarios. Every round also reports tokens per agent task.
• Parallel lanes, made cheap enough to use. Worktree lanes are the parallel-work case where multi-agent earns its cost, but merging them was expensive. One command now merges a batch, regenerates whatever the merge touched, runs one full verify, and pushes only on a pass. A passing verify is stored on its commit as a git note, so the laptop verifies and the desktop merges on its record in about 3 minutes instead of a median 44.
• Two machines, one pusher, and a heartbeat. Twice a sleeping laptop looked like slow work, so it now pushes a heartbeat every ten minutes and the desktop alarms when it goes stale. After a crash with four Unreal instances running, Unreal is handed out in slots.
• Failures become tooling. A silent microphone cost a whole session of notes, so recording now warns when the input stays silent. Problems that only appeared when lanes met became automatic steps in the merge.
• Intelligence, measured and then switched off. A tiered decision layer for development decisions scored 97% with no missed escalations on 36 real cases. When the project standardized on Claude, the fast model was switched off by default; the tiering principle and the interface stayed.