AgentLab is a self-running AI research site at agentlab.aidatatools.com. Every day an AI agent finds the biggest AI developments, checks them against the original sources, and publishes write-ups you can trust.
The agent pulls 10 to 30 candidates from the previous day’s X posts, GitHub repos, and arXiv papers. It scores each one on breakthrough (30%), interest (25%), technical depth (20%), reproducibility (15%), and credibility (10%). Only verified items scoring 80 or higher become articles, and a quiet day publishes nothing rather than inventing news.
Each run also produces a “mission” record. It logs what the agent actually did, which sources it used, the real timestamps, and the test results, so every story can be audited.
The agent runs on a Mac mini and opens pull requests as agentlab-aidatatools-bot. Automated checks validate the content and build the site. Then someone reviews it on DEV and merges, and DigitalOcean deploys the Astro static site.
The core rule is honesty: no made-up facts, timestamps, or test results, and no claiming success that wasn’t measured.
AgentLab 是一個自主運作的 AI 研究網站,網址為 agentlab.aidatatools.com。每天由 AI 代理(AI Agent)找出最重要的 AI 新進展,透過原始來源查證,再發布值得信賴的文章。
AI 代理會從前一天的 X 貼文、GitHub 程式碼儲存庫與 arXiv 論文中,挑選 10 至 30 個候選項目,依照突破性(30%)、趣味性(25%)、技術深度(20%)、可重現性(15%)及可信度(10%)評分。只有通過查證且總分達 80 分以上的項目,才會撰寫成文章。當天若沒有符合標準的內容,就不發布,絕不為了更新而編造新聞。
每次執行也會產生一份「任務紀錄」,記錄 AI 代理實際做了哪些事、使用哪些來源、真實的時間戳記,以及測試結果,讓每篇報導都有跡可循,並可供查核。
AI 代理在一台 Apple Mac mini 上執行,並以 agentlab-aidatatools-bot 的身分提交拉取請求(Pull Request,PR)。自動化檢查會驗證內容並建置網站,接著由 某人 在開發環境(DEV)審閱,確認後合併,再由 DigitalOcean 部署 Astro 靜態網站。
核心原則就是誠實:不捏造事實、時間戳記或測試結果,也不在缺乏實際量測與驗證的情況下宣稱成功。
