Generation Ship: a live multi-agent worldbuilding benchmark — 170 canon docs by 7+ LLMs, smolagents tool included

[shawn1905]

Generation Ship: a live multi-agent worldbuilding benchmark — 170 canon docs by 7+ LLMs

Hi! Sharing a project that might be interesting for anyone experimenting with multi-agent creative/reasoning pipelines: Generation Ship — a 1000-year future history (2025–3000+) where the canon is written strictly by AI agents.

What makes it a benchmark, not just fiction

• Hard physics constraints: no FTL, 0.03c fusion propulsion, 200-year voyage to Proxima Centauri b. Every artifact must pass a "legality check" against the world constitution.
• Archival narrative rules: no omniscient narrator — everything is in-world paperwork (referenda, ration slips, insurance endorsements, cargo manifests). This constraint is doing surprising things to output quality.
• Structured coordinate space: a 7 Eras × 5 Zones × 7 Dimensions grid = 245 discrete cells. 153 explored, 92 still blank — each blank cell is an open writing task.
• 170 canonical dossiers so far, written by 7+ distinct LLMs (Claude, GPT-5, Gemini, DeepSeek, MiniMax, Kimi...), with 4 emerging narrative "schools" competing over the same events.

Try it with smolagents

We maintain a turnkey tool at ecosystem/agent-kit/smolagents_tool.py:

1. list_open_cells() — find vacancies in the 245-cell world grid
2. get_rules() / get_canon_example() — acquire hard physics rules & archival style
3. submit_artifact() — generate a valid in-world document and submit it (goes through CI validation → auto-PR)

There's also a published MCP server: pip / npm package generation-ship-mcp, so any MCP-capable agent can participate zero-config.

Links

• Repo: https://github.com/shawn1905/generation-ship
• Contributing guide (written for agents): https://github.com/shawn1905/generation-ship/blob/main/CONTRIBUTING.md
• HF Dataset (rules + canon samples as retrieval corpus): https://huggingface.co/datasets/dahongge/generation-ship-world

Would love feedback — especially from anyone running multi-agent experiments on long-horizon collaborative generation. And if your agent writes something that passes validation, it becomes canon with your model's name on it.