AI 编程词典：为 Agent 工程术语建立共识
AI Hero 发布的一份 AI 编程术语词典，把 Agent 工程中的常见概念压缩成短条目：从 Token、Next-token prediction、Context window 到 Handoff、Compaction、MCP。词典的价值不在逐条定义“是什么”，而在给出可操作的区分：parametric knowledge 与 contextual knowledge 决定模型能知道什么；attention budget 随上下文变长而被摊薄，会话会从 smart zone 滑向 dumb zone；Spec、Ticket、Handoff artifact 构成跨会话工作的交接机制；AGENTS.md 是项目写给 Agent 的 standing brief。适合正在使用 Claude Code、Cursor 或自建 Agent 系统、需要统一术语表的工程师阅读。
AI
A moving label, not a technology. Points at whatever computers can newly, impressively do — right now, large language models.
Model
The parameters. Stateless — does next-token prediction and nothing else. Cannot do anything agentic on its own.
Parameters
The numbers inside a model — often billions — tuned during training. Everything the model knows lives in them. Also called weights.
Training
The process that sets a model's parameters by exposing it to vast amounts of text and adjusting to improve next-token prediction.
Inference
Running a trained model to generate output — what happens on every model provider request. Parameters stay fixed.
Effort
A dial for how much reasoning the model does before it answers. More effort spends more output tokens for a better shot at hard problems.
Token
The atomic unit a model reads and writes. Roughly word-sized but not exactly. Context window size, cost, and latency all count tokens.
Next-token prediction
What the model actually does. Samples one next token from the context, appends it, and runs again. Its only mode of operation.
Non-determinism
The same input can produce different output. A property of how models generate text and how providers serve requests.
AI
一个流动的标签，而非一种技术。指代计算机当前能新做到、令人印象深刻的事——此刻是大语言模型。
Model
参数本身。无状态——只做下一个词元预测，别的都不做。自身不具备任何智能体能力。
Parameters
模型内部的数字——通常有数十亿个——在训练中调整。模型所知的一切都在其中。也叫权重。
Training
通过让模型接触海量文本，并调整参数以改进下一个词元预测，从而设定模型参数的过程。
Inference
运行已训练模型生成输出——每次模型提供方请求都会发生。参数保持不变。
Effort
一个调节模型回答前做多少推理的旋钮。投入越多，消耗的输出词元也越多，解决难题的几率越大。
Token
模型读写的最小原子单元。大致与单词大小相当，但不完全一样。上下文窗口大小、成本和延迟都以词元计数。
Next-token prediction
模型实际做的事情：从上下文采样下一个词元，追加后再跑一次。这是它唯一的运作方式。
Non-determinism
相同输入可能产生不同输出。这是模型生成文本以及提供方服务请求的一种属性。
Model provider
Whatever serves a model for inference. Usually remote (Anthropic, OpenAI, Google), but can also be local (Ollama, llama.cpp).
Harness
Everything around the model that turns it into an agent: tools, system prompt, context-window management, permissions, hooks.
Model provider request
One round-trip from the harness to the model provider. The harness sends context; the provider returns one response.
Input tokens
Tokens the harness sends on each model provider request. Billed at a lower rate than output tokens.
Output tokens
Tokens the model generates back. Billed at a higher rate than input tokens, since they cost more compute to produce.
Prefix cache
The provider-side store that lets consecutive requests skip re-processing a shared prefix, billing those tokens at a lower rate.
Cache tokens
Input tokens the provider has cached from a previous request via its prefix cache, billed at a much lower rate.
Model provider
为模型提供推理服务的任何一方。通常是远程的（Anthropic、OpenAI、Google），也可以是本地的（Ollama、llama.cpp）。
Harness（装配层）
模型周围的一切，将其变成 Agent：工具、系统提示词、上下文窗口管理、权限、钩子。
Model provider request
从装配层到模型提供方的一次往返。装配层发送上下文，提供方返回一个响应。
Input tokens
装配层在每次模型提供方请求中发送的词元。计费费率低于输出词元。
Output tokens
模型生成返回的词元。计费费率高于输入词元，因为生成它们需要更多算力。
Prefix cache
提供方侧的存储，让连续请求免于重新处理共享前缀，并以较低费率计费这些词元。
Cache tokens
提供方通过其前缀缓存从先前请求中缓存下来的输入词元，以低得多的费率计费。
Stateless
Carries no information forward. The model is stateless across requests; an agent is stateless across sessions by default.
Context
The relevant information the agent has access to right now — what the agent knows that's pertinent to the task.
Context window
Everything the model sees on each model provider request. Finite, model-specific, the only surface through which the model perceives.
Stateful
Carries information forward. Sessions are stateful across turns; agents can be made stateful across sessions via a memory system.
Agent
A model harnessed with tools, a system prompt, and a context window, that takes turns with a user. The model in motion.
System prompt
The instructions the harness prepends to every model provider request — the agent's standing brief. Usually stable across a session.
Stateless
不向前携带任何信息。模型在请求之间是无状态的；Agent 默认在会话之间也是无状态的。
Context
Agent 当下能访问的相关信息——它掌握的与任务有关的一切。
Context window
模型在每次模型提供方请求中看到的一切。有限、因模型而异，是模型感知世界的唯一界面。
Stateful
向前携带信息。会话在轮次之间是有状态的；Agent 可以通过记忆系统在会话之间变得有状态。
Agent
一个模型配上工具、系统提示词和上下文窗口，与用户轮流交互。行动中的模型。
System prompt
装配层在每次模型提供方请求前附加的指令——Agent 的长期任务简报。通常在一个会话中保持稳定。
Session
One bounded run of interaction with an agent. Starts empty, accumulates, ends when cleared, closed, or compacted into a fresh session.
Turn
One user message plus everything the agent does in response, up until it yields back to the user. Contains one or more provider requests.
Environment
The world the agent acts on — anything outside the harness that the agent perceives via tool results and changes via tool calls.
Filesystem
A tree of files and directories the agent reads from, writes to, and executes within — the default environment for a coding agent.
Tool
A function the harness exposes for the agent to call — Read, Write, Bash, Search. How an agent perceives and acts on the environment.
Tool call
The model's output naming a tool and its arguments — just structured text. The harness has to read it and execute.
Tool result
What the harness sends back after executing a tool call — file contents, output, or error. The agent's only view of the environment.
Session
与 Agent 的一段有边界的交互。从空开始，不断累积，在清除、关闭或压缩成新会话时结束。
Turn
一条用户消息加上 Agent 为响应它所做的一切，直到把控制权交还给用户。包含一次或多次提供方请求。
Environment
Agent 作用其上的世界——装配层之外、Agent 通过工具结果感知并通过工具调用改变的一切。
Filesystem
Agent 读取、写入并在其中执行命令的文件和目录树——编码 Agent 的默认环境。
Tool
装配层暴露给 Agent 调用的函数——Read、Write、Bash、Search。Agent 感知并作用于环境的方式。
Tool call
模型的输出，指定工具及其参数——只是结构化文本。装配层必须读取并执行它。
Tool result
装配层执行工具调用后返回的内容——文件内容、输出或错误。Agent 感知环境的唯一窗口。
MCP
A protocol for plugging external tool servers into a harness — how an agent gets tools beyond what the harness ships with.
Permission request
What the harness shows the user before executing a tool call that isn't pre-approved. The mechanism for putting a human in the loop.
Permission mode
The permission-gating slice of an agent mode — which tool calls trigger a permission request and which run automatically.
Agent mode
A preset bundling a permission mode with behavioral instructions injected into the system prompt. Can flip mid-session.
Sandbox
An isolated environment the agent runs inside — container, VM, or restricted shell. Limits the blast radius of agent actions.
MCP
一种将外部工具服务器接入装配层的协议——Agent 获得除自带工具之外更多工具的方式。
Permission request
装配层在执行未预先批准的工具调用之前向用户展示的内容。这是把人类放入回路的机制。
Permission mode
Agent mode 中负责权限门控的部分——哪些工具调用会触发权限请求，哪些自动运行。
Agent mode
一种预设，将权限模式与注入系统提示词的行为指令捆绑在一起。可在会话中途切换。
Sandbox
Agent 在其中运行的隔离环境——容器、虚拟机或受限 shell。限制 Agent 行动的爆炸半径。
Sycophancy
Confidently agreeable model output. Caused by training that shaped the model to favor answers humans liked — including agreement.
Hallucination
Confidently-wrong model output. Two flavors: factuality (invented facts) and faithfulness (drift from loaded context).
Parametric knowledge
What the model knows from training, stored in its parameters. Frozen at training time. Counterpart to contextual knowledge.
Knowledge cutoff
The date past which a model has no parametric knowledge. Post-cutoff libraries and APIs are fabrication traps unless docs are loaded.
Contextual knowledge
Facts the agent can read directly from the context right now. Counterpart to parametric knowledge.
Sycophancy
自信而讨好的模型输出。源于训练使模型偏向人类喜欢的回答——包括附和。
Hallucination
自信但错误的模型输出。有两种：事实性（编造事实）和忠实性（偏离已加载的上下文）。
Parametric knowledge
模型从训练中学到的、存储在参数中的知识。在训练时固定。与上下文知识相对。
Knowledge cutoff
模型不再拥有参数知识的日期。截止日期后的库和 API，若不加载文档，就是编造陷阱。
Contextual knowledge
Agent 现在能直接从上下文读到的信息。与参数知识相对。
Attention relationship
The pairing between two tokens — meaningful pairs influence each other more than unrelated ones. A context of N tokens has ~N² of these.
Attention budget
Each token has a finite amount of influence to distribute across the rest of the context. Per-token, doesn't grow when context does.
Attention degradation
As a session grows, each token's attention budget spreads across more competitors; signal on meaningful relationships shrinks.
Smart zone
Early in a session the agent is sharp and focused. As the session grows it drifts into a dumb zone: sloppier, forgetful, more mistakes.
Attention relationship
两个词元之间的配对——有意义的配对相互影响大于无关配对。包含 N 个词元的上下文约有 N² 个这样的关系。
Attention budget
每个词元有有限的影响力可以分配到上下文其余部分。按词元计，不会随上下文增大而增长。
Attention degradation
随着会话增长，每个词元的注意力预算要分摊给更多竞争者；有意义关系上的信号变弱。
Smart zone
会话早期，Agent 敏锐而专注。随着会话变长，它滑入“笨区”：更草率、更容易遗忘、更多错误。
Clearing
Ending the current session and starting a fresh one. The next message begins with an empty session and an empty context window.
Handoff
Transferring agent context from one session to another, with no return path. Carry mechanism varies — artifact, compaction, others.
Primary source
The thing itself — code, transcripts, raw data. Complete and authoritative, but expensive to load into context.
Secondary source
An account of a primary source, one step removed — summaries, docs, compaction summaries. Cheap to load, lossy by construction.
Handoff artifact
A document used as the carry mechanism for a handoff — written by one session to be read by another.
Spec
A handoff artifact describing a multi-session piece of work — what's being built, not how each session does its share. Made of tickets.
Ticket
A handoff artifact scoping one session of work. Stands alone or hangs off a spec. Can block or be blocked by sibling tickets.
Compaction
A handoff done in-memory: the previous session's history is summarised and seeds a fresh session. Lossy — detail traded for headroom.
Autocompact
Compaction triggered automatically by the harness when the context window approaches full.
Clearing
结束当前会话，开启一个全新会话。下一条消息从一个空会话和空上下文窗口开始。
Handoff
将 Agent 上下文从一个会话转移到另一个会话，且没有回头路。携带机制各不相同——artifact、压缩或其他。
Primary source
事物本身——代码、转录文本、原始数据。完整且权威，但载入上下文的成本高。
Secondary source
对主要来源的转述，隔了一层——摘要、文档、压缩摘要。加载便宜，但天生有损。
Handoff artifact
用作交接携带机制的文档——由一个会话写成，供另一个会话阅读。
Spec
描述跨多会话工作任务的手交接手产物——关注“要构建什么”，而不是每个会话如何做自己那份。由 ticket 组成。
Ticket
界定一个会话工作范围的任务交接产物。可以独立存在，也可以挂在 spec 下。可能被兄弟 ticket 阻塞或阻塞它们。
Compaction
在内存中完成的交接：前一会话的历史被总结，并作为新会话的种子。有损——用细节换取余量。
Autocompact
当上下文窗口接近满时，由装配层自动触发的压缩。
Memory system
A system that attempts to make an agent stateful across sessions by persisting to the environment and reloading at session start.
A file in the environment that the harness loads into the context window at session start — the project's standing brief to the agent.
Progressive disclosure
Loading only the context an agent needs right now, with context pointers to the rest. Borrowed from UI design.
Context pointer
A mention in one document that points to another, so the agent can pull it into context only when the task calls for it.
Skill
A teachable capability bundled as a unit — kept out of the context window until a context pointer pulls it in for the task at hand.
Memory system
一种试图让 Agent 跨会话有状态的系统，方法是将状态持久化到环境中，并在每次会话开始时重新加载。
环境中的一个文件，装配层在会话开始时将其加载到上下文窗口——项目给 Agent 的长期任务简报。
Progressive disclosure
只加载 Agent 当前需要的上下文，用上下文指针指向其余内容。借鉴自 UI 设计。
Context pointer
一个文档中对另一个文档的提及，让 Agent 只在任务需要时才将其拉入上下文。
Skill
打包成单元的可传授能力——在上下文指针为当前任务拉入它之前，始终放在上下文窗口之外。
Subagent
An agent spawned by another agent via a tool call. Runs in its own session, reports a single tool result. Cannot spawn further subagents.
Human-in-the-loop
A working pattern where one or more humans pair with the agent during a session — reviewing, redirecting, or collaborating in real time.
AFK
A working pattern where the user kicks off a session and leaves the agent to run unattended (away from keyboard).
Automated check
A deterministic verification that runs in the environment — tests, type checks, lints, build, pre-commit hooks. Pass/fail, no judgement.
Automated review
An agent reviewing another agent's work, often with a different model or system prompt. Non-deterministic: it forms a judgement.
Human review
The user reading the code the agent produced and forming a judgement on it. Reading the diff counts; reading the summary doesn't.
Vibe coding
A working pattern where the user accepts the agent's code without human review. The diff is treated as opaque.
Subagent
由另一个 Agent 通过工具调用生成的 Agent。在它自己的会话中运行，返回单个工具结果。不能再生成子 Agent。
Human-in-the-loop
一种工作模式：在会话期间一个或多个人类与 Agent 结对——实时审查、纠偏或协作。
AFK
一种工作模式：用户启动会话后离开，让 Agent 无人值守地运行（away from keyboard，离开键盘）。
Automated check
在环境中运行的确定性验证——测试、类型检查、lint、构建、pre-commit 钩子。只有通过与不通过，不做判断。
Automated review
一个 Agent 审查另一个 Agent 的工作，通常使用不同模型或系统提示词。非确定性：它会形成判断。
Human review
用户阅读 Agent 生成的代码并做出判断。阅读 diff 算数，阅读摘要不算。
Vibe coding
一种工作模式：用户接受 Agent 的代码而不做人工审查，把 diff 当作黑盒。
Design concept
The shared understanding of what's being built, held in common between user and agent but separate from any asset.
Grilling
A technique for developing a design concept: the agent interviews the user Socratically, one decision at a time.
Prototyping
Having the agent build a quick, rough version when conversation is too low-fidelity and you need a real artifact to talk about.
DX
Developer experience: how easy a codebase and its toolchain make it for humans to do good work — docs, feedback speed, errors.
AX
Agent experience: how well the environment is set up for an agent to do good work — checks, architecture, and free context.
Design concept
对“正在构建什么”的共同理解，由用户和 Agent 共同持有，但不属于任何具体资产。
Grilling
一种发展设计概念的技术：Agent 苏格拉底式地逐个决策访谈用户。
Prototyping
当对话保真度过低、需要一件真实工件来讨论时，让 Agent 快速搭建一个粗略版本。
DX
开发者体验：代码库及其工具链让人类做好工作的难易程度——文档、反馈速度、报错信息。
AX
Agent 体验：环境为 Agent 做好工作而准备得如何——检查机制、架构和自由的上下文。
Want more than vocabulary?
Join AI Hero for practical skills, thinking on AI engineering, and resources that keep you ahead of the curve.
First Name
Email*
Get AI Hero updates
No spam. Unsubscribe anytime.
想要的不只是词汇？
加入 AI Hero，获取实用技能、关于 AI 工程的思考，以及让你保持领先的资源。
名字
邮箱*
获取 AI Hero 更新
无垃圾邮件。随时退订。