奥特曼小号泄密：OpenAI代码工作100%交给Codex，工程师才揭底Codex“大脑”运行逻辑，碾压Claude架构？

Core Insights - OpenAI has successfully managed to support 800 million users with a single PostgreSQL master database and 50 read replicas, showcasing the efficiency of their architecture over more complex solutions [25][26] - The article reveals the inner workings of Codex's Agent Loop, which coordinates user interactions, model queries, and tool executions to perform meaningful software tasks [5][9][10] Group 1: Codex and Agent Loop - Codex encompasses a range of software agent products, including Codex CLI, Codex Cloud, and Codex VS Code plugins, all built on the same framework [5] - The Agent Loop receives user input, prepares it as a prompt for the model, and queries the model to generate responses, which can include tool calls for further actions [8][9] - The process involves multiple iterations where the agent may call tools and append their outputs to the original prompts, continuing until a final user-facing message is generated [10][12] Group 2: Technical Optimization and Scalability - OpenAI has optimized PostgreSQL to handle over 10 times the load it experienced a year ago, maintaining low latency and high reliability, with 99th percentile latency in the low milliseconds [25] - The company is exploring further performance enhancements by migrating write-intensive workloads to sharded databases like CosmosDB and collaborating with Microsoft Azure for cascading replication [26] - OpenAI's approach emphasizes simplicity and best practices, contrasting with industry trends favoring complex distributed systems [25]