What is WorldSeed?
Don’t build a workflow. Seed a world.
rules + different agents + consequences -> emergence
Define roles, rules, private information, actions, and consequences. Then agents interact until useful artifacts emerge.
You can watch from above, intervene, or step into a character. The same engine can run production rooms, simulations, games, and fictional worlds.
Demo
WorldSeed is scene-agnostic. The same engine runs any world you define.
Scene 1: Autoresearch
You give the system any rough thought or half-formed idea, and a cohort of specialists pursues it. They propose hypotheses, run experiments, peer-review and cite each other’s papers, just like a real research community.
In this run, the goal was to lower val_loss on a 5M GPT trained on TinyStories. In 11 hours:
- 100 hypotheses, 86 experiments, 72 peer-reviewed papers
- val_loss down 24.7%
Every paper is auditable end-to-end: hypothesis, commit, experiment, verified result, citations, reviewer reasoning, forming a search evolution graph.
We observed emergent behaviors such as:
- Role drift. The data specialist stopped finding wins in her own lane early. By the back half of the run she was drafting hypotheses in her teammates’ territory: attention design, second-order optimization. The other two stayed put. Nothing in the config told her to.
-%25E2%2586%2592-blue%3Fstyle%3Dfor-the-badge)
One agent studies a new API. Builder agents create competing demos. Critics reject anything generic. Audience agents judge what feels useful. A curator ships the strongest artifact with its trail of attempts, critiques, and revisions.
Scene 3: AI Layoffs
https://github.com/user-attachments/assets/d43f5d22-1ba8-4483-b720-145b244ddb8c
In an age of AI-driven layoffs, how do people hold on?
One internet company just pulled the trigger: 30% of its workforce, gone.
Those being laid off have to “distill” their expertise into an AI Skill before they leave. Distill honestly, or leave a backdoor in the Skill?
Those who stay face the same deadlines, higher KPIs, and twice the workload. Grind it out, or quietly plan the exit?
Four people in this office, each with their own play running:
- The PM everyone loves working with. Who’s he really trashing the moment the door closes?
- The architect walking out at month’s end. Severance didn’t land. What gets buried in the Skill he hands over?
- The team lead who demands honest data from everyone. Can her own “AI productivity” numbers survive a closer look?
- The QA nobody remembers is there. Those bugs in his private folder: evidence, or ammunition?
Try it locally
Scene 4: Teahouse Espionage
Same engine. Different YAML. Completely different world.
Four spies, one teahouse. Who’s really working for whom?
A classic espionage drama in miniature. Agents trade secrets over tea, protect their covers, and try to read each other before they get read.
Getting Started
Prerequisites: Python 3.11+, Node.js 18+, uv
git clone https://github.com/AIScientists-Dev/WorldSeed && cd WorldSeed
uv sync --extra dm
cd frontend && npm install && npm run build && cd ..
cp .env.example .env
# Add your API key (any LiteLLM provider: OpenAI, Anthropic, Ollama, etc.)
uv run worldseed play configs/ai_layoffs.yaml
Open the dashboard at http://localhost:8000. Three ways to experience it:
- Watch: observe all agents from above, including their inner state.
- Intervene: whisper privately to any agent, nudge the story.
- Play: step into a character and play alongside the AI.
Every run is different. Past runs are preserved and replayable.
Choose the runtime guide you need:
How It Works
WorldSeed runs on a tick loop over a world you declared in YAML. A tick is one beat of the world’s clock, like a heartbeat that advances the world one step at a time. Each tick: every agent perceives its own filtered slice, proposes an action, and the engine resolves it. Predictable outcomes follow the rules you declared; uncertain ones go to an AI referee. Effects apply, the world advances, the next tick begins.
Setup (once, in YAML):
- Any world, in one YAML file. Declare entities, rules, physics, and per-character perception; the engine has zero hardcoded domain knowledge.
Runtime (every tick):
- Asymmetric information by design. Perception rules filter the world per character. Three agents in the same room hold three completely different pictures of what’s happening.
- Deterministic rules where you can, AI judgment where you can’t. Predictable actions resolve instantly via the in-YAML rule engine (DSL); uncertain ones go to an LLM-based Dungeon Master (DM) that returns structured effects, not free prose.
- Effects land, the world ticks on. State mutates, consequences cascade, the next tick begins. Slow or offline agents don’t freeze it, and every change is logged for replay.
Plug-in points:
For the full runtime plumbing (endpoints, tick scheduling, consequences, inbox delivery), see Architecture. See a real scene YAML in configs/teahouse.yaml or the full schema in Scene Config Spec.
Create Your World
Describe your world in a prompt, let AI generate the YAML, then hand-craft whichever pieces you want more control over: a character’s secret, a specific action’s rule, a perception filter, a DM hint.
Generate with AI:
/create-world "An AI tool pilot lab where builders create competing demos, critics reject generic outputs, and a curator ships the strongest artifact"
The create-world skill produces both YAML scene config and UI config, validated and ready to run.
Hand-craft any feature:
The output is plain YAML. You can edit any entity, action, rule, character profile, or perception filter directly. Study the built-in examples (teahouse.yaml, ai_layoffs.yaml) to see how features are declared.
Full spec: Scene Config · UI Config · DSL
Validate and run:
uv run worldseed validate configs/my_scene.yaml
uv run worldseed play configs/my_scene.yaml
Once launched, each scene auto-renders room cards, character portraits, and a narrator voice you pick (storyteller / noir / intel briefing / gossip):
| Room Cards |
Character Portraits |
 |
 |
Development
Follow Getting Started above, then:
uv sync --all-extras
# Tests
uv run pytest tests/ -q # all
uv run pytest tests/unit/ -q # fast, no IO
uv run pytest tests/e2e/ -v # real server
uv run pytest tests/scenarios/ -q # scene-agnostic
# Lint, format, type-check
uv run ruff check --fix src/ tests/
uv run ruff format src/ tests/
uv run mypy src/
Join the discussion or ask for help:
MIT. See LICENSE.
For anyone building multi-agent worlds. Run the bundled scenes, or create your own.
评论