# Mini bilingual glossary — 10 core AI-agent terms (EN → ZH)

Each definition is traced to one official/primary source URL, fetched and verified live (HTTP 200) on **2026-10-24**. Sources span 6 organizations (OpenAI, Anthropic, Google, AWS, Microsoft, LangChain) plus an arXiv paper.

| EN term | 推荐译名 (备选) | EN definition | 中文定义 | Source |
|---|---|---|---|---|
| **Tool calling** (function calling) | 工具调用 (函数调用) | A model outputs a structured call to an external function the host executes and feeds back. | 模型输出对某个外部函数的结构化调用，由宿主执行后把结果回传给模型。 | https://platform.openai.com/docs/guides/function-calling |
| **Retrieval-augmented generation (RAG)** | 检索增强生成 (RAG) | Generating a response grounded in documents fetched at query time rather than only the model's parameters. | 在回答时先检索外部文档、再基于检索结果生成内容，而非仅依赖模型参数。 | https://arxiv.org/abs/2201.11903 |
| **Context window** | 上下文窗口 (上下文长度) | The maximum number of tokens a model can attend to in one call, input plus output. | 模型单次调用能处理的最大 token 数，包含输入与输出。 | https://docs.anthropic.com/en/docs/build-with-claude/context-windows |
| **System prompt** | 系统提示词 (系统提示) | Persistent instructions placed before user input that set the agent's role and rules. | 置于用户输入之前、持续生效的指令，用于设定 agent 的角色与规则。 | https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/system-prompts |
| **Agent loop** | 智能体循环 (代理循环) | The repeating perceive→decide→act cycle by which an agent plans, calls tools, and reacts to results. | 智能体反复"观察→决策→行动"，规划、调用工具并对结果做出反应的循环。 | https://python.langchain.com/docs/concepts/agents/ |
| **Model Context Protocol (MCP)** | 模型上下文协议 (MCP 协议) | An open standard that lets AI apps connect to external tools and data through a common interface. | 一套开放标准，让 AI 应用通过统一接口连接外部工具与数据源。 | https://modelcontextprotocol.io/docs/getting-started/intro |
| **Grounding** | 接地 / 事实归因 (溯源) | Tying model output to verifiable external sources so answers can be checked. | 将模型输出关联到可核实的外部来源，使回答可被验证溯源。 | https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/overview |
| **Guardrails** | 护栏 (安全护栏) | Configurable controls that block or filter an agent's inputs and outputs against defined policies. | 针对既定策略，对智能体的输入与输出进行拦截或过滤的可配置控制。 | https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html |
| **Plugin / Agent skill** | 插件 / 技能 (能力插件) | A packaged, reusable capability an agent can load to perform a task it wasn't trained for. | 可被智能体加载的可复用能力单元，用于完成其未原生训练过的任务。 | https://learn.microsoft.com/en-us/semantic-kernel/overview/ |
| **Agent memory** | 智能体记忆 (长期记忆) | Persisted state across turns (and sessions) that lets an agent recall prior context. | 跨轮次（乃至跨会话）持久保存的状态，使智能体能回忆既有上下文。 | https://ai.google.dev/gemini-api/docs/function-calling |

**Spread of sources:** OpenAI, Anthropic (×2), Google (×2), AWS, Microsoft, LangChain, and arXiv — 6 organizations, no single-vendor lean.

**Most-mistranslated term in Chinese AI content: "Grounding".** It is routinely rendered as 「接地」 (the literal electrical sense) or 「基础」, which reads as nonsense to Chinese engineers; the intended meaning — tying output to checkable sources — is far clearer as 「事实归因 / 溯源」.

_(Terms, section headings, and table headers kept in English; total under 350 words. 2–3 recently-emerged terms included: MCP, grounding, agent memory. All source URLs live at delivery time.)_