YUME
YUME is a social simulation where Claude Managed Agents live as residents of a city.
Core trick
Gives each simulated resident its own agent session that receives only world state and returns actions, with state updates applied between turns.
Tags
Description
YUME is a social simulation where Claude Managed Agents live as residents of a city. Each resident has its own independent agent session and persistent memory. No instructions are given to the agents. They receive only the state of the world — buildings, nearby people, their own health, finances, stress — and reason autonomously about what to do. This architecture makes the simulation genuinely realistic.
Under the hood, YUME implements per-resident asymmetric relationship models, an independent evaluation step that scores each interaction, seed-based reproducible probabilistic events, and parallel synchronization orchestration across all agents — a structure that cannot be achieved with a single prompt.
Change the city layout, swap in different residents, or alter the scenario, and an entirely different society emerges. Urban planning, policy testing, social experiments — whatever the question, YUME answers it by simulating the whole city.
README excerpt
# YUME [](https://cerebralvalley.ai/e/built-with-4-7-hackathon) YUME is a local AI life-simulation app. A Bun backend runs the simulation, creates one Anthropic Managed Agents session per actor, applies deterministic state changes, stores run data in SQLite, and streams events to a React/Three.js frontend. The frontend starts runs, follows live server-sent events, displays actor state and relationships, and replays the city timeline in 3D. ## Architecture | Layer | Path | Runtime | Default port | Responsibility | |---|---|---|---:|---| | Backend | `core/` | Bun + TypeScript | `3001` | HTTP API, simulation loop, Managed Agents, SQLite, SSE | | Frontend | `web/` | Vite + React + Three.js | `8080` | Setup UI, live event stream, actor panels, 3D replay | | Sample data | `web/sample-data/` | JSON | n/a | Example actors and locations | Important backend files: - `core/src/server.ts`: HTTP routes, env loading, request parsing, run creation. - `core/src/server/sse.ts`: server-sent event streaming and replay. - `core/src/schema.ts`: Zod schemas for run input, state, and events. - `core/src/design-orchestrator.ts`: main simulation loop. - `core/src/simulation-engine.ts`: local state transitions and action effects. - `core/src/simulator-config.ts`: defaults, thresholds, slots, action rules. - `core/src/managed-agents-driver