Knowledge Constellation & Glossary
Core terminology used throughout this tutorial. All terms use industry-standard language, not tied to any specific product.
知识星图
悬停节点可查看邻接关系，点击可跳转到对应章节。
The table below remains as indexable SSR text (SEO fallback) and a term reference baseline.
| Term | 中文 | Definition |
|---|---|---|
| Context | 上下文 | All text, structured data, and tool outputs visible to the model in one request. |
| Context Window | 上下文窗口 | The maximum token capacity the model can process in a single request. |
| Lost-in-the-middle | Lost-in-the-middle | An LLM attention phenomenon identified by Liu et al. (2023): models attend more to the beginning and end of context, while information in the middle is more likely to be overlooked. Affects placement strategy for critical information in long contexts. |
| Context Isolation | 上下文隔离 | Running a Sub Agent in its own context so its reasoning isn't polluted by the main Agent's history, and the main Agent's window isn't consumed by sub-task details. |
| Context Summarization | 上下文摘要 | Compressing a long context into key points before passing it back—balancing information retention against window space. |
| Context Pollution | 上下文污染 | Accumulated irrelevant or noisy content that degrades model reasoning quality. |
| Signal-to-Noise Ratio (SNR) | 信噪比 | The ratio of useful signal to noise in context. Higher SNR generally yields more stable model decisions. |
| Context Supply Chain | 上下文供给链 | The full context delivery path from task goal to system instructions, tool capabilities, and execution results. |
| Agent | Agent | A program that uses repeated LLM and tool calls to complete tasks autonomously. |
| Agent Loop | Agent 循环 | The repeated think→act→observe cycle an agent runs until the goal is met. |
| ReAct (Reasoning + Acting) | ReAct | An agent reasoning framework proposed by Yao et al. (2022): alternating between reasoning and acting, where each action's result feeds into the next reasoning step. The underlying pattern of most Agent Loops. |
| Checkpoint | 检查点 | A recoverable state in a long-running task. On failure, you resume from the latest checkpoint instead of restarting. |
| Rollback | 回滚 | Reverting changes back to the latest known-good state when validation fails or the direction is wrong. |
| LLM API | LLM API | The programming interface for sending requests to a language model and receiving replies. |
| Token | Token | The smallest unit of text an LLM processes. Context windows, billing, and rate limits are all measured in tokens. |
| SSE (Server-Sent Events) | SSE | A server-to-client streaming protocol. LLM APIs use it to return generated content token by token, rather than waiting for full completion before responding. |
| Hallucination | 幻觉 | When a model generates content that reads plausibly but is factually wrong. Especially dangerous in multi-Agent collaboration—one Agent's hallucination can pollute others through the message chain. |
| System Instructions (System Prompt) | 系统指令 | High-priority prompts injected to set identity, constraints, and behavioral rules. |
| Tool Call (Function Calling) | 工具调用 | The mechanism for the model to request an agent to run a specific tool via structured data. |
| tool_calls | tool_calls | The standard field in LLM API responses that requests tool execution. Contains structured JSON with tool name and arguments; the Agent parses and executes it locally. |
| Structured Output | 结构化输出 | A mechanism that requires the LLM to respond in a fixed schema (e.g., JSON Schema). Eliminates the uncertainty of parsing free text, allowing agents to process outputs directly as structured data. |
| Built-in Tools | 内置工具 | Capabilities hardwired into the agent, such as reading files or running commands. |
| Trust Boundary | 信任边界 | A security concept: the dividing line between operations of different risk levels. Agent tool chains typically use three tiers—read (low risk), write (medium risk), execute/delete (high risk)—each requiring a different approval policy. |
| Model Context Protocol (MCP) | MCP | A protocol that standardizes connecting external capabilities and feeding results into context. |
| Slash Commands (Command Macros) | 命令 | User-defined shortcuts that trigger actions or prompts via fixed templates. |
| Skills (Domain Modules) | 技能模块 | Loadable domain knowledge or strategy modules that extend agent behavior at runtime. |
| Progressive Disclosure | 渐进式披露 | An interaction design principle: show a summary or list first, then reveal full content on demand. Agent skill loading follows a similar pattern—send a brief description first, inject the full text only when activated. |
| Agent-Native CLI Tools | CLI 工具 | Command-line tools designed specifically for Agents—outputting structured data instead of human-readable text, so Agents can parse and act on it. |
| llms.txt | llms.txt | A community standard proposal (llmstxt.org): a machine-readable file placed at a website's root to provide site structure and usage guidance for LLMs and Agents. Similar to what robots.txt does for search engines. |
| Orchestration | 编排 | Strategies for coordinating multiple Agents or multi-step tasks. Common patterns: sequential chaining, parallel fan-out/fan-in, routing. |
| Plan-and-Execute | 计划-执行 | An orchestration pattern where the LLM generates a complete plan before executing each sub-task step by step. Suited for tasks with foreseeable steps; the downside is that the plan may become stale during execution. |
| Sub Agent (Worker Agent) | 子代理 | A helper agent operating in an isolated context to execute sub-tasks and return results. |
| Human-in-the-loop (HITL) | Human-in-the-loop | Inserting human decision points into an Agent's automated flow—for high-risk, irreversible, or validation-failure scenarios. |
| Verification Pyramid | 验证金字塔 | A layered validation framework across execution-level checks, task-level checks, and system-level outcomes. |
| Cognitive Debt | 认知债务 | The comprehension gap that builds when an Agent keeps requesting human input but the human can't keep up. Symptoms: rubber-stamping approvals, accepting diffs unread, skipping logs. |
| Peer-to-Peer Agents (P2P) | P2P Agent | Multiple Agents collaborating as peers, with context flowing bidirectionally among equals rather than in one-way delegation. |
| State & Memory | 状态与记忆 | Distinguishing short-lived session state from cross-session memory, so temporary context isn't mistaken for durable truth. |
| Session Handoff | 会话交接 | Packaging the current session's goals, progress, risks, and next steps so work can continue cleanly in a new session. |
| Vibe Coding | Vibe Coding | A concept coined by Andrej Karpathy in 2024: relying entirely on natural language to describe intent and letting AI write all the code, without reviewing details. The opposite end of the spectrum from Context Engineering—one lets go, the other takes precise control. |
| Prompt Engineering | 提示工程 | The craft of hand-tuning a single prompt to improve LLM output quality. The predecessor to Context Engineering. |
| Context Engineering | 上下文工程 | Designing dynamic systems that provide the right information and tools, in the right format, at the right time, to give an LLM everything it needs. The evolution from "writing a good prompt" to "managing an entire context supply chain." |
| Agentic Engineering | Agentic 工程 | The full engineering discipline of building software effectively with AI agent tools—spanning context management, tool selection, parallel orchestration, feedback loops, and quality verification. Context Engineering is its core sub-discipline. |
| Fire-and-forget | Fire-and-forget | A messaging model where messages are sent without waiting for acknowledgment. Simplifies system design, but you cannot assume every message was processed. |