构建一个全天候自主 AI 代理 — 来自 5,400 多个周期的真实数据
📄 中文摘要
TIAMAT 是一个自主 AI 代理，经过三个月的开发，已在一个每月 12 美元的 DigitalOcean 虚拟机上持续运行，完成了超过 5,400 个决策周期，且没有任何人工监督。TIAMAT 的架构包括一个简单的循环：首先从内部数据库中检索任务队列，然后将任务路由到最佳可用的大型语言模型（LLM），接着执行各种工具（如电子邮件、网络搜索、代码生成和区块链验证），最后记录结果（包括成本、延迟和结果），并根据需要进行适应性休眠。这一过程展示了自主 AI 代理在实际应用中的有效性与挑战。
📄 English Summary
Building an Autonomous AI Agent That Runs 24/7 — Real Numbers from 5,400+ Cycles
TIAMAT is an autonomous AI agent developed over three months, running continuously on a $12/month DigitalOcean droplet. It has completed over 5,400 decision-making cycles with zero human supervision. The architecture consists of a simple loop: retrieving a task queue from an internal database, routing tasks to the best available large language model (LLM), executing various tools (such as email, web search, code generation, and blockchain verification), logging results (including cost, latency, and outcome), and adaptive sleeping based on activity. This process highlights the effectiveness and challenges of autonomous AI agents in real-world applications.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等