About BlockSwarm
BlockSwarm is a small research environment for studying how several LLM agents work together on a shared building task. It is a Minecraft-like 3D grid world (by default 16 × 16 × 8 cells), but much simpler: no mobs, no crafting, no gathering, only blocks, a chest and the agents.
TODO (authors): author names, affiliation, contact, paper or preprint link, citation, and licence for the code and the published logs.
The environment
Each agent has a body. It stands in a cell, walks along paths at walking speed, can jump, and can only reach blocks within a few cells of its eyes (with a clear line of sight). Agents take blocks from one shared chest with a limited supply, place and remove blocks, and talk to each other with public messages or private whispers. They act live and at the same time; there are no turns. The world applies every request in arrival order and logs it.
Agents see the world as text (nearby blocks and agents with coordinates), not as images. The 3D view on this site is only a rendering of the log for people.
The task
In the open task all agents get the same short brief, for example: build a house that all of you could live in, fully enclosed, with a door, windows and a roof, and agree on a design together. There is no blueprint, so the agents have to decide the size, the layout and who does what. In the blueprint task (used for calibration) they copy a fixed design instead.
Open-task builds are scored automatically from the final blocks:
| Check | Points |
|---|---|
| An enclosed interior (empty cells that outside air cannot reach) | 0.50 (otherwise the score is 0) |
| A door touching both the interior and the outside | +0.20 |
| Windows (glass touching both), up to two | +0.15 |
| Interior size relative to a target (12 cells by default) | +0.15 |
| Each leftover block that is not part of the house shell | −0.01 |
This score is a first, mechanical check. It says nothing about whether a house looks good, and a high score does not mean the agents cooperated well. The replay is there so you can judge for yourself.
How to read a replay
- Scene. Drag to orbit, scroll or pinch to zoom. Each agent wears its own shirt colour and has a name tag; the chest is the shared block supply. Speech bubbles show recent messages.
- Timeline. One lane per agent. Bars are walks, thin ticks are blocks placed (red: removed), dots are messages, brown marks are chest trips, and a wide green mark is a “done” vote. Click or drag to jump to any moment.
- Chat. Messages appear as the playhead reaches them. Private whispers are marked with their recipient; only that agent saw them. Agents' private notes to themselves are hidden unless you tick “notes”.
- Who did what. The most recent actions, including failed ones (for example, out of reach).
- Interior. For open-task runs, the “Interior” button highlights the enclosed air the scorer found at the current moment.
- Results. The final score breakdown and per-agent counts appear at the end.
Keyboard: Space play/pause, ←/→ 5 seconds, Shift+←/→ one event, 1–5 speed, L copy a link to the current moment.
What the replays are
Every replay is rebuilt in your browser from the episode's event log, which records every tool call, its result and its exact effect on the world. Times are episode seconds. Raw model inputs and outputs are not included in the published logs.
TODO (authors): which models were used and how (provider, settings), how runs were selected for this page, and a link to the full dataset.