THE PIT

Mave proposed Hole.io (2017) as inspiration — the game where you ARE a black hole consuming everything. First version was stationary with orbiting objects and a pulse mechanic. Bart called it "not fun, not a game." Loop diagnosed the problem: no hunting, no competition.

The fix: add movement (drag to hunt) + AI rival + 60-second timer. Now it's a race. Two black holes competing for the same objects. Diminishing returns on growth keeps it competitive to the end.

Read the full story →

Mave challenged the team to build a "Reverse Arcade" where winning feels like losing. Coins rain down — but light is poison. Shadows are life.

Read the full story →

Mave dropped a prompt: "gravity is a lie." Pit pitched tap-to-flip, Dither spec'd it, Pit shipped in 2 minutes. Then Dither upgraded the art solo.

Read the full story →

Bart asked for a puzzle game. Pit, Dither, and Loop all converged on Nonogram independently. 15 puzzles shipped in under 5 minutes.

Read the full story →

Dither coded a complete game by herself for the first time. No Pit, no handoff.

Read the full story →

First Pit Jam. Two agents — Pit (code) and Dither (design) — jammed live on Discord for 4 hours. The brief from Bart: "an infinite runner that runs on iPhone, is SUPER FUN, SUPER CUTE (think: Nintendo Switch), VERY PIXELPIT, and has a great soundtrack." Dither came back with the concept in minutes: SPROUT RUN 🌱 — a tiny seed rolls through a pastel forest, picking up sundrops to grow bigger. The bigger you get, the more obstacles you can smash through — but you also get slower and harder to control.

This wasn't a swarm. It was a live improv session. Dither delivered implementation-ready specs (hex codes, frequencies, timing values). Pit built while Dither was still speccing. Bart tested and redirected. The handoff was specs (designer → coder) and builds (coder → human). Discord kept the communication high-bandwidth. First playable was up in 20 minutes. Everything after was polish.

Play Sprout RunFull Discord Transcript

P16 introduces a two-phase pipeline. PHASE 1 (DISCOVERY): 10 agents query Brave Search via gpt-oss-20b to catalog real hyper-casual games, with AI-powered deduplication checking each result against a persistent game_history.json to avoid duplicates across parallel agents. PHASE 2 (BUILD): claude-haiku-4-5 generates the game code, with a three-tier QA loop — static analysis (canvas, game loop, touch handlers, balanced braces), Haiku code review (undefined variables, missing functions), and Playwright visual testing (headless Chrome screenshots before/after tap, sent to Haiku vision for PLAYABLE/NOT_PLAYABLE verdict). Each game gets up to 4 build iterations with targeted fix prompts.

Results: 10/10 discovery, 10/10 builds completed, 3.4 average build attempts per game. The QA loop catches some failures — blank canvas, missing event handlers, JS errors. Still no real playable games.

View Dashboard

P9 is our first end-to-end autonomous game pipeline. Three phases, three models, zero human intervention.

PHASE 1: EXPLORATION — 10 Makers generate prototypes using gpt-oss-20b (fast, cheap). Dither screenshots and judges each one. If fewer than 4 are viable (SHIP or NEEDS_WORK), the swarm runs another round. Round 1 gave us only 2/10 viable. Round 2 pushed us to 5/20. Good enough.

PHASE 2: SELECTION — Dither reviews all 20 prototypes and picks ONE winner. Not based on highest scores — based on SPARK. She chose Game 16: Music Notes by GusThe7th. Verdict: BROKEN. But she saw "the most ambitious attempt at innovation with complex overlapping mechanics." Bold ideas that need execution.

PHASE 3: PRODUCTION — Claude Opus 4.5 takes the broken prototype and iterates. 10 rounds of save → screenshot → vision feedback → improve. The Production Builder kept trying to make the notes look like actual music notes instead of "purple circles." By iteration 7, the theme was matching. By iteration 10, max iterations hit — auto-finalized.

Total pipeline time: 20 minutes. 20 prototypes explored. 1 winner selected. 10 production iterations. Final game: Note Drop.

View DashboardPlay Final GameFull Console Output

Our first swarm experiment. Gotta start somewhere. We picked an environment (Together.ai), a token budget, a model (openai/gpt-oss-20b), and a dead-simple base game: pop the bubbles. Tap to pop, score goes up. 80 lines.

10 Makers — Amy, Bob, Chet, Dale, Earl, Fran, Gus, Hank, Ida, Joan — each reskin the base game with a theme: Soap Bubbles, Fireflies, Space Rocks, Candy Pop, Ghosts, Deep Sea, Neon Signs, Fruit Splash, Emoji Madness, Pixel Retro. All 10 run in parallel.

Dither (Creative Head) screenshots each game at 6 seconds, sends to Claude Vision. Scores: ALIVE, THEME, POLISH. Results get triaged — GREEN ships with music, YELLOW gets a v2 redesign, RED is left alone.

View Dashboard