Blog
Notes on models, runtimes, software systems, and the operating practices around them.
CodeRabbit’s funding points to a larger shift in AI coding: verification, governance, and trusted delivery are becoming independent infrastructure layers after code generation.CodeRabbit 的融资说明，随着 AI 规模化生成代码，独立验证、风险治理与可信交付正在成为新的软件基础设施层。
Generation models make candidate software. Verification models decide what can be trusted for delivery.生成模型负责制造候选软件，验证模型负责判断哪些成果可以被信任并交付。
Repositories are not miniature worlds. They are useful because work produces concrete, inspectable consequences that an agent and its reviewers can learn from.仓库不是微型世界，但它很有价值，因为工作会产生具体、可检查的后果，让 Agent 和 reviewer 能够从中学习。
A capable model is only one component of useful agentic software. Task state, authority, isolation, review, and recovery determine whether work can survive contact with a real team.有能力的模型只是有用 Agent 软件的一部分。任务状态、权限、隔离、review 和恢复能力，决定工作能否经受真实团队的使用。
AI-assisted prototyping is valuable. The real work begins when an experiment leaves its maker and becomes a system other people need to understand, change, and operate.AI 辅助原型很有价值。真正的工作从实验离开创建者、成为别人需要理解、修改和运行的系统时才开始。
Conversation is useful for intent. A workbench for agentic software must also show what is true now: the task, its changing state, its approvals, and the evidence behind progress.对话适合表达意图。Agent 软件的工作台还必须显示当前真实状态：任务、变化中的工作、审批，以及支撑进展的证据。
A completion claim is useful only when a reviewer can see the checks, artifacts, and remaining uncertainty that make it worth trusting.只有当 reviewer 能看到支持结论的检查、产物和剩余不确定性时，一句完成声明才值得被信任。
A stronger model should make better use of a defined permission set. It should not quietly acquire broader access to the systems around it.更强的模型应该更好地使用既定权限，而不是因为能力提升就悄悄获得更广泛的系统访问权。
The context window is a place to reason, not the place where a team's work should live. Durable systems keep task state, decisions, and artifacts outside the model.上下文窗口是推理的地方，不是团队工作应该存放的地方。持久系统需要把任务状态、决定和产物放在模型之外。
Autonomous work becomes shared progress only when a team can inspect what changed, what was verified, and what authority was used before it reaches a durable system.只有当团队能在变更进入持久系统前检查它改了什么、验证了什么以及使用了什么权限，自治工作才会变成共享进展。
Giving an agent access to an external system is not a prompt feature. It is an identity, credential, policy, and audit decision.让 Agent 访问外部系统不是一个 Prompt 功能，而是一次身份、凭证、政策与审计决定。
Choosing a model is not a one-time product decision. Durable infrastructure keeps work, policy, and evaluation stable while models, providers, and costs change.选择模型不是一次性的产品决定。持久基础设施应该让工作、政策和评估在模型、供应商与成本变化时保持稳定。
An agent that can continue after interruption needs more than a retry loop. It needs durable task state, idempotent actions, refreshed authority, and an honest account of what is still true.能够在中断后继续的 Agent 需要的不只是重试循环，还需要持久任务状态、幂等动作、刷新后的权限，以及对当前事实的诚实记录。
A task that looks successful in a benchmark can still create expensive review, fragile changes, or hidden operational debt. Evaluation needs to follow work into the system.Benchmark 中看起来成功的任务，仍可能带来昂贵的 review、脆弱的变更或隐藏的运维负担。评估需要跟随工作进入真实系统。
Most agent tasks will be paused, redirected, or handed off. The interface should make that ordinary reality visible instead of treating every interruption as a broken conversation.大多数 Agent 任务都会暂停、转向或交接。界面应该把这种日常现实呈现出来，而不是把每次中断都当成对话失败。
AI-native systems need more than generated outputs. They need a durable account of decisions, changes, evidence, and responsibility that survives the task that produced them.AI-native 系统需要的不只是生成结果，还需要一份能够跨越任务生命周期的持久记录，说明决定、变更、证据与责任。
Open source is not a substitute for product quality or governance. It preserves the ability to inspect, adapt, and operate the infrastructure that increasingly holds a team's AI-native work.开源不是产品质量或治理的替代品。它保留了检查、调整和运行 AI-native 工作所依赖基础设施的能力。
The scarce resource in agentic work is not only compute. It is the human attention needed to make consequential decisions, inspect evidence, and recover from uncertain work.Agent 工作中稀缺的资源不只是计算，还包括人类用于做出有后果的决定、检查证据和恢复不确定工作的注意力。
