I built AgentPages, an experiment in running AI agents entirely inside GitHub. https://github.com/idorozin/AgentPages The idea is simple: instead of just hosting a static site on GitHub Pages, the site is maintained by an AI agent that lives in the repo. You define what you're interested in (topics, goals, style), and the agent periodically researches those topics and updates the site. Everything runs inside GitHub: A GitHub Agentic Workflow (gh-aw) runs on a cron schedule The agent performs web research (using Tavily search) It updates its knowledge base It edits the Astro source files The site is rebuilt and deployed to GitHub Pages The whole thing is transparent because everything happens in the repo: user/ files you control (research topics, requests, feedback) agent/ the agent’s knowledge and memory website/src/ the site the agent edits docs/ the built static site served by GitHub Pages You can steer the agent by editing files like: user/profile.md what the agent should research user/feedback.md give feedback on its work user/requests/ ask it to investigate specific topics Every run creates a pull request with the changes (which can optionally be auto-merged). A couple example sites powered by AgentPages: https://idorozin.me/agentpages-site — project documentation maintained by the agent https://idorozin.me/iran-monitor — a live conflict tracker maintained by an agent This is very much an MVP / experiment. I mostly wanted to explore the idea of “GitHub-native agents” where the agent: lives in the repo runs via GitHub Actions publishes to GitHub Pages can be steered through files Curious what people think about this model. |