用一个 PostgreSQL 主库和 50 个只读副本，就顶住了 ChatGPT 上的 8 亿用户！ 近日，OpenAI 的工程师们不仅爆出了这一惊人消息，还直接把 Codex 的"大脑"给扒了个精光。在 OpenAI 官方工程博客主页，OpenAI 工程师、Technical Staff 成员 Michael Bolin 发布了一篇文章，以"揭秘 Codex 智能体循环"为题，深入揭秘了 Codex CLI 的核心框架：智能体循环（Agent Loop），并详细讲 解了 Codex 在查询模型时如何构建和管理其上下文，以及适用于所有基于 Responses API 构建智能体循环的实用注意事项和最佳实践。 这些消息传出后，在 Hacker News 等技术论坛及社交平台上获得了高度关注。"看似平淡的技术最终会胜出。OpenAI 正在证明，优秀的架构远胜于花哨的 工具。" 值得一提的是，有网友透露，前不久 Anthropic 的一位工程师称"他们用于 Claude Code UI 的架构糟糕且效率低下"。而就在刚刚，X 上出现一条爆料： Codex 已接管 OpenAI 100% 的代码编写工作。 Agent ...

整理 | 华卫 Codex 的"大脑"揭秘 "每个人工智能智能体的核心都是 Agent Loop，负责协调用户、模型以及模型调用以执行有意义的软 件工作的工具之间的交互。" 据介绍，在 OpenAI 内部，"Codex"涵盖了一系列软件智能体产品，包括 Codex CLI、Codex Cloud 和 Codex VS Code 插件，而支撑它们的框架和执行逻辑是同一个。 用一个 PostgreSQL 主库和 50 个只读副本，就顶住了 ChatGPT 上的 8 亿用户！ 近日，OpenAI 的工程师们不仅爆出了这一惊人消息，还直接把 Codex 的"大脑"给扒了个精光。在 OpenAI 官方工程博客主页，OpenAI 工程师、Technical Staff 成员 Michael Bolin 发布了一篇文章， 以"揭秘 Codex 智能体循环"为题，深入揭秘了 Codex CLI 的核心框架：智能体循环（Agent Loop），并详细讲解了 Codex 在查询模型时如何构建和管理其上下文，以及适用于所有基于 Responses API 构建智能体循环的实用注意事项和最佳实践。 这些消息传出后，在 Hacker ...

Core Insights - Anthropic's Claude Code has gained significant attention in the AI programming community, with developers praising its capabilities as a future-oriented AI assistant [1] - OpenAI has revealed two major updates: the Agent Loop architecture and the PostgreSQL database's ability to support 800 million users [1][24] Group 1: Agent Loop Architecture - The Agent Loop acts as a "conductor," integrating user intent, model brain, and execution tools into a cohesive system [6] - It operates through a cycle of observation, thinking, action, and feedback, allowing AI to function independently rather than just answering questions [8] - The process involves constructing a detailed prompt, model inference, tool calls, result feedback, and looping until the task is completed [10][12][22] Group 2: Key Technologies in Agent Loop - OpenAI has introduced two optimizations to address the challenges of Agent development: Prompt Caching and Compaction [13][21] - Prompt Caching reduces costs by avoiding the need to resend unchanged parts of the conversation history, lowering the cost growth from quadratic to linear [15][17] - Compaction allows the model to summarize long conversations into a shorter format, preserving essential understanding while managing context limitations [21][23] Group 3: PostgreSQL Database Performance - OpenAI's PostgreSQL database supports 800 million ChatGPT users with a single primary node and 50 read replicas, showcasing its robust infrastructure [24][25] - Key technologies include PgBouncer for connection pooling, cache locking mechanisms to prevent cache misses, and cross-regional replication to distribute read requests [27][30] - The architecture emphasizes read-write separation and optimization of read paths, crucial for handling high volumes of read requests [26][28] Group 4: Competitive Landscape - Claude Code's end-to-end development experience poses a direct threat to OpenAI's Codex CLI, pushing OpenAI to enhance its agent architecture [37][41] - OpenAI's updates signal its strong engineering capabilities and infrastructure, indicating a competitive edge in the AI tools market [38][39] - The ongoing competition between Claude Code and OpenAI's Codex is expected to drive further innovation, ultimately benefiting developers [41]