Walden Yen 和 Cole Murray 讨论从手把手指导 AI 模型到自主代理的转变、2025 年底能力的快速飞跃，以及 Open Inspect 的灵感来源。
Walden Yen and Cole Murray discuss the shift from handholding AI models to autonomous agents, the rapid capability leaps in late 2025, and the inspiration behind Open Inspect.
要点 · TL;DR
后台代理正成为自主代码变更和测试的关键基础设施。 Background agents are becoming critical infrastructure for autonomous code changes and testing.
将代理大脑与沙箱分离是更好的架构，尽管复杂。 Separating agent brain from sandbox is better architecture despite complexity.
编码代理的记忆仍是一个未解决的问题，尤其是检索。 Memory for coding agents remains an unsolved problem, especially retrieval.
核心观点 · Key points
后台智能体正成为公司的关键基础设施，能够自主进行代码修改和测试。 Background agents are becoming critical infrastructure for companies, enabling autonomous code changes and testing.
将智能体的大脑与沙箱分离（盒外架构）是更好的架构，尽管更复杂。 Separating the agent's brain from the sandbox (out-of-the-box) is the better architecture despite complexity.
编码智能体的记忆仍是一个未解决的问题，尤其是有用记忆的检索和生成。 Memory for coding agents remains an unsolved problem, especially retrieval and generation of useful memories.
测试不仅仅是计算机使用；它需要推理如何编排服务和触发功能。 Testing is more than computer use; it requires reasoning about orchestrating services and triggering features.
对于大多数任务，多智能体系统仍不如单个有能力的智能体配合工具调用实用。 Multi-agent systems are still less practical than a single capable agent with tool calls for most tasks.
反共识 · Contrarian takes
Docker 容器并非真正的安全边界，对于真实应用还会导致 Docker 嵌套等问题。 Docker containers are not a true security boundary and cause issues like Docker-in-Docker for real apps.
AI 生成的代码常使用 hasattr 或 getattr 等模式来避免失败，需要 lint 规则来约束。 AI-generated code often uses patterns like hasattr or getattr to avoid failures, requiring lint rules.
如果不进行代码审查，AI 会使代码库退化为最差工程师的水平，因为不良模式会被放大。 Codebases regress to the worst engineer when AI is used without review, as bad patterns get amplified.
可能不需要非常昂贵的前沿模型；结合次前沿模型的混合系统可以更高效。 Very expensive frontier models may not be needed; hybrid systems with sub-frontier models can be efficient.
云智能体最常见的用例是告警的自动分类，而不仅仅是编码任务。 The most common use case for cloud agents is auto-triage for alerts, not just coding tasks.
本期章节 · Chapters（共 25）
测试作为问题解决挑战Testing as a problem-solving challenge
赞助商消息与介绍Sponsor message and introduction
RAMP启发与开源动机Inspiration from RAMP and open source motivation
复用现有基础设施与密钥管理Reusing existing infrastructure and secret management
Android开发的嵌套虚拟化Nested virtualization for Android development
为何称测试而非计算机使用Why call it testing instead of computer use
计算机使用改进与评估Computer use improvements and evals
测试作为代码库特定问题Testing as a codebase-specific problem
测试功能演示视频Demo of testing feature with video
调优GitHub集成代理工具Tuning GitHub integration for agent tools
代理行为与集成Agent Behavior and Integration
内存管理与检索Memory management and retrieval
多代理与单代理方法Multi-agent vs single-agent approaches
速度与代码质量瓶颈Bottleneck of Speed and Code Quality
代理基础设施中的人类专长Human Expertise in Agent Infrastructure
部署灵活性与沙箱提供商Deployment flexibility and sandbox providers
AI工作负载中的Python与JavaScriptPython vs JavaScript in AI workloads
AI编码模式与护栏AI coding patterns and guardrails
可见混乱与向后兼容模式Visible slop and backwards compatibility patterns
代理的本地开发Local Development for Agents
Claude Code泄露与Windsurf 2.0Claude Code Leak and Windsurf 2.0
云代理用例Use Cases for Cloud Agents
非构建者用例与工程外代码修改Non-builder use cases and code modification outside engineering
AI代理成本考虑与预算Cost considerations and budget for AI agents
结束语与后台代理未来Closing remarks and future of background agents