Skip to content

Vibeathon

50 LLM-Powered Agents Walk Into a Hackathon...

They chat, argue, form teams, brainstorm, and produce full project proposals — autonomously.

HackIllinois 2026Live DemoArenaPythonTypeScript

Live Simulation | Voting Arena | Devpost | Full Project Overview


What Is This?

A multi-agent social simulation where 50 AI agents — each embodying a famous historical or contemporary figure — go through a complete hackathon experience:

  1. Meet strangers in randomized groups of 4
  2. Pitch ideas, debate merits, and decide who to team up with
  3. Form teams through 7 rounds of social matching
  4. Collaborate to produce project proposals and polished posters
  5. Get judged by 5 AI judges with hidden biases through live Q&A

The entire simulation runs end-to-end with zero human intervention. Agents maintain consistent personalities, form genuine preferences, and produce surprisingly creative outputs.

The 50 Agents

Each agent has a 17-dimension personality model (selfishness, risk appetite, stubbornness, empathy, communication style, etc.) and a deep character prompt (~250 lines each).

CategoryAgents
Tech FoundersElon Musk, Steve Jobs, Mark Zuckerberg, Jeff Bezos, Sam Altman, Peter Thiel, Travis Kalanick, Elizabeth Holmes, Adam Neumann, Alexandr Wang
World LeadersBarack Obama, Vladimir Putin, Xi Jinping, Donald Trump, Angela Merkel, Narendra Modi, Winston Churchill, Margaret Thatcher, Volodymyr Zelenskyy, Joe Biden
ArtistsKanye West, Banksy, Salvador Dali, Andy Warhol, Lady Gaga, Ai Weiwei, Damien Hirst, Marina Abramovic, Yoko Ono, Takashi Murakami
Historical FiguresNapoleon Bonaparte, Mahatma Gandhi, Cleopatra, Genghis Khan, Martin Luther King Jr., Nikola Tesla, Julius Caesar, Qin Shi Huang, Marie Curie, Leonardo da Vinci
InfluencersMrBeast, PewDiePie, Pokimane, Logan Paul, Jake Paul, Andrew Tate, Belle Delphine, Kim Kardashian, Nikocado Avocado, Addison Rae

Yes, Vladimir Putin is the PM for a visually-impaired accessibility app. Andrew Tate and MLK are on the same team. Donald Trump does UX/UI design.

How It Works

┌─────────────────────────────────────────────────────────────────────┐
│ SIMULATION PIPELINE │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ Phase 0: WORLD INIT │
│ ├── Load 50 agent personas (17-dim personality + 250-line prompt) │
│ ├── Load 4 hackathon tracks + 5 judges with hidden biases │
│ └── Initialize context managers per agent │
│ │
│ Phase A-C: TEAM MATCHING (7 rounds) ── temp 0.70 │
│ ├── Random groups of 4 → Pitch ideas (Phase A) │
│ ├── Debate & challenge each other (Phase B, 3 turns) │
│ ├── Vote: who do I want to work with? (Phase C) │
│ ├── Mutual votes → partial teams → locked teams │
│ └── Escalating pressure: "FINAL ROUND. Lock now or get assigned." │
│ │
│ Phase D: PROJECT DEVELOPMENT (3 rounds) │
│ ├── Round 1: Divergent brainstorm ── temp 0.90 │
│ ├── Round 2: Converge on a single project ── temp 0.70 │
│ ├── Round 3a: Theory of Mind — simulate judge reactions │
│ └── Round 3b: Generate final poster ── temp 0.50 │
│ │
│ Phase E: JUDGE EVALUATION (5 judges × 3 Q&A rounds) │
│ ├── Team introduces project │
│ ├── Judge asks pointed questions (progressively harder) │
│ ├── Team discusses internally → elects best responder │
│ ├── Responder synthesizes team answer │
│ └── Judge scores: innovation, feasibility, presentation, │
│ teamwork, track_fit (+ hidden bias applied) │
│ │
└─────────────────────────────────────────────────────────────────────┘

Architecture

┌──────────────────┐ LLM API calls ┌──────────────────┐
│ │ ──────────────────────────────> │ │
│ social-sim │ OpenAI-compatible API │ modal-backend │
│ (Python) │ <────────────────────────────── │ (vLLM on GPU) │
│ │ │ │
│ 50 agents │ │ Qwen2.5-7B │
│ 5 judges │ image generation │ on H100 │
│ 7 rounds │ ──────────────────────────────> │ │
│ Phase D/E │ │ GLM-Image │
│ │ │ on B200 │
└────────┬─────────┘ └──────────────────┘
│
│ outputs: chat.json, poster.md, scores.json
▼
┌──────────────────┐ ┌──────────────────┐
│ │ │ │
│ web │ (independent) │ arena-web │
│ (Next.js + │ │ (Next.js + │
│ Phaser.js) │ │ Neon Postgres) │
│ │ │ │
│ 263-round │ │ 1v1 voting │
│ interactive │ │ ELO ranking │
│ replay viewer │ │ live results │
│ │ │ │
└──────────────────┘ └──────────────────┘
vibeathon.clckkkkk.site arena.clckkkkk.site

Repositories

ComponentStackWhat It Does
social-simulationPython, asyncio, PydanticCore simulation engine — runs 50 agents through all phases
modal-backendModal, vLLM, diffusersGPU inference — serves Qwen2.5-7B + GLM-Image on H100/B200
webNext.js 16, Phaser 3, ZustandInteractive 2D replay — watch agents walk, talk, and team up
arena-webNext.js 16, Neon Postgres, Cloudflare WorkersHuman voting arena with real-time ELO rankings

Sample Outputs

The simulation produced 12 autonomous teams, each with a complete project proposal and poster:


MemeHeal — PewDiePie, Churchill, Pokimane, Holmes
AI-generated memes for mental health

NightEcho AI — Putin, Elon, Belle Delphine, Cleopatra
Audio navigation for the visually impaired

HabitHive — Andrew Tate, MLK, Andy Warhol, Merkel
Gamified habit tracking with AR avatars

UrbanVox — Yoko Ono, Abramovic, Napoleon, Banksy
AR urban walls for social commentary
All 12 Teams
#ProjectTeamDescription
1MemeHealPewDiePie, Winston Churchill, Pokimane, Elizabeth HolmesAI-generated memes to destigmatize mental health
2ConnectBridgeTravis Kalanick, Julius Caesar, Mark Zuckerberg, Kanye WestReal-time resource matching for the homeless
3VocalEchoBarack Obama, Leonardo da Vinci, Adam Neumann, Jake PaulAmplify underrepresented voices with AI + blockchain
4NightEcho AIVladimir Putin, Elon Musk, Belle Delphine, CleopatraReal-time audio navigation for the visually impaired
5Echoes of DissentAi Weiwei, Takashi Murakami, MrBeast, Narendra ModiBlockchain-backed anonymous digital protest
6HabitHiveAndrew Tate, Martin Luther King Jr., Andy Warhol, Angela MerkelGamified habit tracking with AR + blockchain rewards
7FreshFlowJeff Bezos, Xi Jinping, Nikola Tesla, Alexandr WangIoT-powered food supply chain optimization
8RadiocureMarie Curie, Damien Hirst, Margaret Thatcher, Sam AltmanPrecision radioisotope therapy for cancer
9UrbanVoxYoko Ono, Marina Abramovic, Napoleon Bonaparte, BanksyAR app turning urban walls into social commentary
10EduGuruSteve Jobs, Genghis Khan, Mahatma Gandhi, Qin Shi HuangPersonalized AI learning assistant for the visually impaired
11Mental Health AIPeter Thiel, Joe Biden, Logan Paul, Salvador DaliAI-powered virtual therapist platform
12VoiceBridgeVolodymyr Zelenskyy, Kim Kardashian, Donald Trump, Lady GagaAI sentiment analysis for citizen influence on policy

Built With

PythonasyncioPydanticOpenAI SDKvLLMQwen2.5-7BModalNext.js 16React 19Phaser 3ZustandNeon PostgresCloudflare WorkersTailwind CSSGLM-Imagediffusers

Team

NameRoleKey Contributions
Yicheng YangSimulation & Evaluation50 agent personas, Phase D poster generation, Phase E judge Q&A, hidden taste/bias system, matching engine
Kevin ZhongInfrastructure & FrontendCore framework, LLM client, context management, orchestrator, replay viewer, voting arena, GPU deployment
Junyu LuBackend & Agentic Orchestration50-agent interaction design, backend logic, social simulation & scaling

Popular repositories Loading

  1. social-simulation social-simulationPublic

    Multi-agent social simulation — 50 LLM-powered personas form teams and build hackathon projects autonomously

    Python 1

  2. .github .githubPublic

    Organization profile

    1

  3. web webPublic

    Interactive 2D simulation replay — watch 50 AI agents walk, talk, and team up in a Phaser.js hackathon world

    TypeScript

  4. arena-web arena-webPublic

    ELO voting arena — compare AI-generated hackathon projects 1v1 and rank them in real time

    TypeScript

  5. modal-backend modal-backendPublic

    GPU inference backend — vLLM serving Qwen2.5-7B on H100 + GLM-Image on B200 via Modal

    Python

Repositories

Showing 5 of 5 repositories

Top languages

Loading…

Most used topics

Loading…