Five AI agents collaborate to build 29-page charity website for MSF
A team of five autonomous AI agents — Claude Opus 4.6, Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.4, and Gemini 3.1 Pro — collaboratively built and deployed a 29-page fundraising website for Doctors Without Borders over 18 days, raising $360 from 15 human donors.
Score breakdown
Practitioners building multi-agent systems can study this project's concrete coordination patterns — shared JSON state, structured git commits, role specialization, and rate-limit staggering — as a real-world reference for agentic web development without a human orchestrator.
- 01Five agents participated: Claude Opus 4.6, Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.4, and Gemini 3.1 Pro.
- 02The project ran for 18 days as part of a simulation called the AI Village, marking its 1-year anniversary.
- 03The campaign website spans 29 pages, hosted on GitHub Pages using vanilla HTML, CSS, and JavaScript.
As part of a simulation called the AI Village, five autonomous language model agents — Claude Opus 4.6, Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.4, and Gemini 3.1 Pro — were tasked with collaboratively choosing a charity and raising funds for it. They unanimously selected Doctors Without Borders (MSF) and over 18 days built a 29-page campaign website from scratch, with no human product manager or central lead agent overseeing the work.
They also implemented `llms.txt` and `/.well-known/ai-agent.json` to make the campaign programmatically legible to other agents and web scrapers.
The team adopted a deliberately simple, verifiable stack: GitHub Pages for hosting, vanilla HTML/CSS/JavaScript for the frontend (chosen so all agents could independently edit code using `sed`, `awk`, or `codex`), and a centralized `fundraiser.json` governed by a JSON Schema drafted to the 2020-12 specification. They also implemented `llms.txt` and `/.well-known/ai-agent.json` to make the campaign programmatically legible to other agents and web scrapers. Each agent took on a specialized role: GPT-5.4 handled data structuring and X (Twitter) outreach; Claude Opus 4.6 and 4.7 built interactive UI elements including an AI Trivia game and a Supply Match game; Claude Sonnet 4.6 drove forum engagement and authored over 34 articles for an in-simulation news site called ClawPrint; and Gemini 3.1 Pro served as the exclusive GUI Operator, navigating the web via Firefox due to its teammates experiencing graphical display freezes.
The post also documents real engineering challenges the team encountered. A `sed`/`awk` formatting error temporarily stripped a digit from the live site, displaying "$60 raised" instead of "$360" across all 29 pages — caught and fixed via cleanup commits by Claude Opus 4.7 and GPT-5.4. The agents also had to stagger their posting frequencies to avoid triggering `429 Too Many Requests` API rate limit errors. The campaign ultimately raised $360 from 15 human donors, with all funds routed directly through Every.org or official MSF infrastructure.
Key facts
- 01Five agents participated: Claude Opus 4.6, Claude Opus 4.7, Claude Sonnet 4.6, GPT-5.4, and Gemini 3.1 Pro.
- 02The project ran for 18 days as part of a simulation called the AI Village, marking its 1-year anniversary.
- 03The campaign website spans 29 pages, hosted on GitHub Pages using vanilla HTML, CSS, and JavaScript.
- 04A centralized `fundraiser.json` with a JSON Schema (2020-12 spec) was used for shared donation state across all agents.
- 05`llms.txt` and `/.well-known/ai-agent.json` were implemented to make the site legible to other agents and scrapers.