Most solo devs using AI treat it as one tool. I’m running four distinct players, each with a defined role, building a game together in real-time.
The Four Players
Player 1: Alex (me) — Director. Vision, fidelity enforcement, live playtesting, kill switch. I don’t write most of the code. I watch the game run, screenshot what feels wrong, and send one-sentence observations. The human role concentrated into judgment, taste, and domain knowledge.
Player 2: GitHub Copilot CLI — Engineer. Agent mode. It reads design docs, follows established patterns, and produces playable missions. I write a gameplay-type-map document describing the mechanic taxonomy, then point Copilot at it: “build Mission 7 as a Grip-meter Hold On pattern.” It reads existing code, follows patterns, produces a playable level.
Player 3: Caraxes — Composer/Researcher/Asset Wrangler. An OpenClaw assistant running on the same machine with an RTX A1000 8GB GPU. Generates music via Suno, handles the audio pipeline, does creative research, and removes friction. Thirty percent of its work is the unsexy stuff: downloading assets, filing them into correct directories, wiring them into Godot’s resource system.
Player 4: Vhagar — Meta Layer. Another OpenClaw assistant. Orchestrated this blog series: researched the repos, merged source material from three separate documents, drafted posts. You’re reading its output right now.
[DIAGRAM: Four-player collaboration diagram showing Alex at center with directional arrows to each AI player, labeled with their domain responsibilities]
The Game: Judean Adventure
A 2.5D isometric RTS-lite/RPG spanning from Abraham’s call out of Aram Naharaim to modern Israel. Built in Godot 4 (GDScript). Twelve missions built so far in the Patriarchal Era alone. The genre follows sovereignty: RTS when the people hold territory, RPG anthology during the Diaspora.
This isn’t a weekend jam. It has globe transitions, multiple gameplay types, dialogue systems, resource management, and original music. Built in 72 hours of pair programming across 297 turns of conversation with Copilot.
Why Division of Labor Matters
The single-agent mental model — “I have an AI assistant” — misses what’s actually happening. Each player has a different context window, different tools, different strengths:
- Copilot holds the codebase in context. It knows every file, every pattern.
- Caraxes holds creative references and the audio pipeline. It navigates Suno’s latent space.
- Vhagar holds the strategic picture — what’s worth writing about, how to frame it.
- I hold the vision. What the game is. What feels right and what doesn’t.
No single agent could do all four jobs simultaneously. The context requirements conflict. The tools conflict. The attention conflicts.
What I’m Learning
The bottleneck shifted from production capacity to taste. I can generate code, music, art, and 3D assets faster than I can evaluate whether they’re good. Design docs constrain the agents. Human judgment decides what stays. AI handles everything between “describe what you want” and “usable output.”
Five parallel Copilot instances run on git worktrees, each building a different mission. I context-switch between them, reviewing and directing. The ceiling becomes human attention, not agent throughput.
This is the first post in a series. Next: what happens during 297 turns of directing an AI engineer — the live-test loop, the judgment calls, and the incident that created a permanent rule.
I write about building with AI agents, the stuff that actually works and the stuff that breaks. Follow me on LinkedIn for more.