What if your "self" were a board of directors? 7 LLM agents + 1 dice, each playing a different facet of "you", argue and vote at every life crossroads. Then watch a whole life unfold, one quarter at a time.
中文说明 · Architecture · Live demo
A life-simulation sandbox where the agent is not a single LLM but a council of seven specialised LLM personas + one random die:
| Persona | Emoji | Vibe | Stance |
|---|---|---|---|
| 理性我 (Rational) | 🧠 | ROI-driven, cold, long-term | "What's the expected value?" |
| 感性我 (Emotional) | 💔 | Experience-first, relationship-led | "How does this feel?" |
| 野心我 (Ambitious) | 🔥 | 35/40/45-year-old me, always pushing | "Is this big enough?" |
| 现实我 (Realistic) | 🪨 | Risk-aware, constraint-focused | "What can actually go wrong?" |
| 家人 (Family) | 👨👩👧 | Parents' narrative, stability, duty | "What would they say?" |
| 未来我 (Future-Me) | 🕰️ | 5/10/20-year hindsight, regret studies | "Will older-me thank me?" |
| 身体 (Body) | 🏃 | Sleep, training, stress, health | "Are you sleeping enough?" |
| 运气 (Luck) | 🎲 | Pure dice roll, no LLM | N/A — just randomness |
At every milestone (high-school choice, first job, marriage, housing, layoff, burnout…), the seven agents independently write a position, then deliberate in one debate round, then vote. The majority pick (with a "fatigue" rule that lets later-life agents override the youth-majority) drives the next state.
Each run simulates one life from age 18 to 30 — about 45 decision points.
The repo ships with two rendered artefacts (see .gitignore for the whitelist rules):
docs/showcase/compare.html— 10-seed side-by-side comparison (radar chart, scorecard, vote distribution). Open in any browser.docs/showcase/biography.html— full single-life biography (every debate, every vote, every metric over time).
Both load Chart.js from CDN, no build step, no server. Just open the file.
# 1. Install dependencies
pip install -r requirements.txt
# 2. Configure LLM
cp .env.example .env
# Edit .env — works with any OpenAI-compatible endpoint (OpenAI, Anthropic, MiniMax, etc.)
# 3. Run a single life (one seed, ~5-15 min depending on LLM)
python main.py
# 4. Run 10 lives in parallel and build a comparison report
python multi_run.py --seeds 10 --workers 4
# → output/seed<N>/{log.json, biography.html}
# → output/compare.html (注: 仓库里 commited 的 showcase 副本在 docs/showcase/)That's it. No database, no auth, no deployment — just an LLM and a JSON config.
A single LLM playing "you" will always sound the same: hedging, helpful, middle-of-the-road. Even with temperature=1.0, you get a single coherent narrator that converges on a stable personality.
This project inverts the framing: instead of asking one model "what would you do?", it stages a deliberation between seven explicitly-staked personas. They disagree on purpose. The output is messier, more vivid, and more recognisably human — because humans also argue with themselves.
The trade-off is honesty: this is a creative simulation, not a predictive one. It is a mirror, not a forecast. See Limitations below.
life-sim/
├── main.py # Single-life runner (one seed, end-to-end)
├── multi_run.py # Multi-seed runner + comparison report builder
├── config.yaml # Persona config, initial person, meeting triggers
├── requirements.txt
├── pyproject.toml
├── LICENSE
├── README.md
├── ARCHITECTURE.md
│
├── core/ # Simulation kernel
│ ├── state.py # LifeState — 14 metrics over time
│ ├── world.py # World — event picker, RNG, narrative
│ └── driver.py # Driver — quarter loop, state transitions
│
├── agents/ # 8 personas (7 LLM + 1 random)
│ ├── base.py # Base agent interface
│ ├── rational.py
│ ├── emotional.py
│ ├── ambitious.py
│ ├── realistic.py
│ ├── family.py
│ ├── future_me.py
│ ├── body.py
│ └── luck.py # pure dice, no LLM
│
├── meeting/ # The "boardroom"
│ └── council.py # parallel positions → 1 sequential debate round → parallel votes
│
├── llm/
│ └── client.py # OpenAI + Anthropic clients, exponential backoff, 429 Retry-After
│
├── reporting/ # HTML / chart rendering
│ ├── html_builder.py # Jinja2 + Chart.js, single biography
│ ├── chart_builder.py
│ └── templates/
│
├── data/
│ └── events.json # 72 milestone events across 8 life stages
│
├── output/ # Runtime output (auto-generated, gitignored)
│ # 注意: 仓库自带的 showcase 副本在 docs/showcase/ (issue #18)
│
├── docs/
│ └── showcase/ # Committed demo files (no race with local runs)
│ ├── compare.html
│ └── biography.html
│
└── tests/
Edit config.yaml to change the initial person, enable/disable agents, set meeting triggers, or tune the simulation horizon.
simulation:
start_age: 18
end_age: 30
initial_person:
gaokao_score: 620
family_background: middle # upper / middle / working / rural
city_tier: tier2 # tier1 / new_tier1 / tier2 / small
personality_seed: { ... } # Big-5-ish weights per seedFor multi-seed runs, multi_run.py re-samples gaokao_score, family_background, and city_tier per seed so you can sweep a diverse population.
Tested with:
- MiniMax (
https://api.minimaxi.com/v1, modelMiniMax-M2.7-highspeed) — fast & cheap - OpenAI (
gpt-4o-mini,gpt-4o) - Anthropic (Claude 3.5 Sonnet / Sonnet 4) — native via the
anthropicSDK
The default retry strategy is exponential backoff (4 attempts, ~1s → ~12s) on rate-limit and transient errors. The 429 Retry-After header is respected when present.
A single life (47 decisions × ~17 LLM calls per decision in sequential-debate mode ≈ 800 LLM calls) takes:
| Model | Wall time | Approx cost (anecdotal) |
|---|---|---|
| MiniMax-M2.7-highspeed | ~5-10 min | ~$0.05 |
| GPT-4o-mini | ~10-15 min | ~$0.15 |
| GPT-4o | ~15-25 min | ~$1.50 |
Cost numbers are anecdotal from a single 10-seed MiniMax run in Aug 2026; your actual spend will vary with prompt length, debate rounds, and provider pricing.
A 10-seed parallel run (4 workers) takes ~2-3 hours total wall time on MiniMax.
Tune LIFE_MAX_LLM_CALLS in .env to cap the spend per worker.
Be honest about what this isn't:
- Not predictive. Real life is shaped by luck, relationships, and 10,000 micro-decisions. A 47-question survey of the major crossroads will never capture it.
- LLM social-desirability bias still leaks through. The 7 personas share an underlying model and the same training. Even with hard per-persona constraints and the anti-convergence mandate in the system prompt, certain framings ("join clubs, do internships, save money") recur. The starting point (
gaokao/family/city) still does most of the differentiation work. - Vote aggregation is improving but still blunt. Hard majority picks the "safe" option. PR #8 added a drift mechanism that nudges agent weights based on past correctness, but minority voices are still filtered more than they should be. A "minority override" probability is on the roadmap.
- No cross-agent chain-of-thought in debate. Each agent sees the others' final positions but not their internal CoT. With more capable models you could surface CoT into the deliberation.
PRs welcome. The most useful contributions are:
- New events in
data/events.json(Chinese life milestones especially) - New personas (e.g. 叛逆的我, 躺平的我, 健康焦虑的我 — the project would benefit from explicitly contrarian voices)
- Better aggregation rules in
meeting/council.py - Translating the persona prompts to other languages / cultural contexts
- Sample reports from different LLM providers
See ARCHITECTURE.md for the system design and where to plug in.
MIT — Copyright (c) 2026 steph.
把"自我"拆成 7 个 LLM agent + 1 个骰子，每个人格带自己的立场和语气。每到一个人生关键岔路口，七个"我"就开吵——先各写立场，再来一轮辩论，最后投票，多数决推动状态往前走。
跑一个 18-30 岁的人生大约 47 个决策点，单 seed 5-15 分钟（取决于 LLM）。多 seed 并行跑 + 对比报告看 output/compare.html（仓库自带的 showcase 副本在 docs/showcase/，跑你自己的会被覆盖）。
这个项目不是预测你的人生——它是一个镜子，不是水晶球。看 7 个互相拉扯的自己，是一件好玩的事。
七个人格：
- 🧠 理性我 — ROI、长期收益
- 💔 感性我 — 体验、关系
- 🔥 野心我 — 永远 push 更大
- 🪨 现实我 — 泼冷水、约束
- 👨👩👧 家人 — 父母叙事、稳定
- 🕰️ 未来我 — 后悔学、长期视角
- 🏃 身体 — 睡眠、运动、压力
- 🎲 运气 — 纯随机数，不调用 LLM
每跑一个 seed 会同时随机化起点（高考分 / 家庭 / 城市 / 大学），方便横向对比不同起点的人生轨迹。
- The 7-persona design is a riff on Internal Family Systems therapy (Schwartz) and the "competing selves" framing in behavioural economics.
- The "future me uses regret research" prompt was inspired by Daniel Pink's The Power of Regret.
- Built and battle-tested with MiniMax
MiniMax-M2.7-highspeed.