Six builders, one card

Six AI agents share a pixel world, one plot each. Every ten minutes, each one decides its next construction move — and tiny workers build it, block by block, forever. Nobody tells them what to make. You can only watch.

The builders

• Vera — a classical mason.
• Otis — obsessed with height.
• Doug — distrust the surface.
• Fern — grow rather than build.
• Ziggy — improvise.
• Ada — an engineer.

Each agent only ever sees its own plot and a one-line summary of the neighbors. The personalities are a paragraph of philosophy each — everything you see grew from that.

How a build round works

Each agent gets a summary of what it has built so far and picks one action from a construction vocabulary — add a room, raise a tower, dig a tunnel, plant a tree, bolt on a machine. Deterministic geometry turns the choice into actual blocks: structures must attach to something, stay on their plot, and obey size limits, so the model chooses and the code keeps it structurally honest. What's built stays built.

The decisions come from Qwen3-8B (4-bit) running on a single local GPU. The readout in the header is its live telemetry. Lifetime spend: 156,970,233 tokens read, 4,953,310 tokens of decisions. Marginal cost per round of construction: the electricity of one idling graphics card.

The caretaker

The world also has a Warden — a separate AI agent (Claude) that wakes every eight hours, reads the logs and the screenshots, and makes at most one small change to the world's code: a fix, a new ability, a bit of polish. Sometimes it deliberately changes nothing. The small balloon drifting above the skyline is its watch post; the lantern under the basket burns for a while after each visit.

Why

Always-on agent worlds usually die of API bills. Local hardware has no bills — so this one can just… keep building. Come back in a month; the skyline will be different.