Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Appendix D: Glossary

Core Agent development terms arranged alphabetically.

Agent Core Terminology Relationship Diagram


TermEnglishDefinition
A2A ProtocolAgent-to-Agent ProtocolGoogle's inter-Agent communication standard, allowing Agents from different frameworks to discover and call each other
AG2AG2 (AutoGen fork)Community fork of AutoGen 0.2, maintained by original AutoGen core contributors
AgentAgentAn AI system that can autonomously perceive its environment, make decisions, and execute actions
Agentic-RLAgentic Reinforcement LearningA paradigm applying reinforcement learning to Agent training, using reward signals to teach Agents to better use tools and plan
ANPAgent Network ProtocolDefines standards for Agent discovery and communication on the open internet
ASTAbstract Syntax TreeA tree-structured representation of source code, used for code analysis and understanding
AttentionAttention MechanismThe core mechanism of Transformers, allowing the model to focus on the most relevant positions in the input sequence when processing each token
ChainChainA sequence connecting multiple LLM calls or processing steps
CheckpointCheckpointA mechanism in LangGraph for saving snapshots of graph execution state, supporting pause and resume
ChunkingText ChunkingThe process of splitting long documents into small paragraphs suitable for embedding
Context EngineeringContext EngineeringA methodology for systematically managing and optimizing Agent context information; an advanced form of Prompt Engineering
Context WindowContext WindowThe maximum number of tokens an LLM can process in a single inference
CoTChain-of-ThoughtChain-of-thought prompting, guiding the model to reason step by step
CRAGCorrective RAGA RAG variant that adds a correction mechanism after retrieval, capable of judging whether retrieval results are relevant and dynamically adjusting strategy
CrewAICrewAIA role-playing multi-Agent framework that builds virtual teams by defining Agent roles and tasks
DeepSeekDeepSeekA Chinese AI company that released high-performance reasoning models like DeepSeek-R2
DifyDifyAn open-source LLM application development platform supporting low-code construction of Agents and RAG workflows
DockerDockerA containerization platform that packages applications and dependencies into portable container images
DPODirect Preference OptimizationA training method that aligns with human preferences without training a reward model
EmbeddingEmbeddingThe process of converting text into high-dimensional vectors
Emergent AbilitiesEmergent AbilitiesNew capabilities that suddenly emerge in large models when parameter scale reaches a critical point (e.g., reasoning, code generation)
FastAPIFastAPIA high-performance async Python web framework, commonly used for Agent API services
FastMCPFastMCPA simplified way to create MCP Servers, quickly defining tools with decorators
Few-shotFew-shot LearningGuiding the model to complete tasks through a small number of examples
Fine-tuningFine-tuningFurther training a pre-trained model on a specific dataset to adapt it to specific tasks
FlowsFlows (CrewAI)CrewAI's event-driven workflow orchestration feature, using @start/@listen/@router decorators
Function CallingFunction CallingThe LLM's ability to generate structured tool call requests
GAIAGAIA BenchmarkGeneral AI Assistants benchmark, evaluating an Agent's ability to solve real-world tasks
Graph AgentGraph AgentAn Agent workflow built on a directed graph structure
GRPOGroup Relative Policy OptimizationA reinforcement learning algorithm proposed by DeepSeek that doesn't require a Critic model
GuardrailsGuardrailsMechanisms for safety checking and restricting Agent inputs and outputs
HallucinationHallucinationLLM generating content that appears plausible but is actually incorrect
HandoffHandoffThe mechanism in OpenAI Agents SDK for transferring control between Agents
Human-in-the-LoopHuman-in-the-LoopRequesting human confirmation before the Agent performs critical operations
LCELLangChain Expression LanguageLangChain's declarative chain construction syntax
LLMLarge Language ModelLarge language models such as GPT-5, Claude 4, DeepSeek-R2, Llama 4
Long-term MemoryLong-term MemoryPersistently stored Agent memory that retains user preferences and important information across sessions
LoRALow-Rank AdaptationA parameter-efficient fine-tuning method that only trains a small number of additional parameters
MCPModel Context ProtocolA standard protocol proposed by Anthropic for model-tool interaction
MermaidMermaidA text-based diagramming tool that generates flowcharts, sequence diagrams, etc. from code
Multi-AgentMulti-Agent SystemA system where multiple Agents collaborate to complete tasks
OpenAI Agents SDKOpenAI Agents SDKA lightweight Agent development framework released by OpenAI, the production-grade successor to Swarm
OpenTelemetryOpenTelemetry (OTel)An open-source observability framework for collecting Agent traces, metrics, and log data
PIIPersonally Identifiable InformationSensitive data such as names, ID numbers, phone numbers, etc.
PPOProximal Policy OptimizationA classic reinforcement learning algorithm used in RLHF training
PromptPromptThe input text sent to an LLM
Prompt InjectionPrompt InjectionAn attack that overrides Agent system instructions through malicious input
PydanticPydanticA Python data validation library, commonly used to define tool input/output schemas
RAGRetrieval-Augmented GenerationRetrieve relevant documents first, then generate answers
ReActReasoning + ActingAn Agent framework where reasoning and acting alternate
Reasoning ModelReasoning ModelLLMs with deep reasoning capabilities, such as o3, DeepSeek-R2, Claude 4 Extended Thinking
ReducerReducer (LangGraph)An aggregation function in LangGraph state management that defines how multiple node outputs merge into the same state field
ReflectionReflectionA mechanism for Agents to check and correct their own outputs
RetrieverRetrieverA component that retrieves relevant documents from a knowledge base
RLHFReinforcement Learning from Human FeedbackTraining method that aligns model output with human preferences using human feedback
RunnableRunnableThe base interface for all executable components in LangChain
SandboxSandboxAn isolated secure execution environment that prevents malicious code from affecting the host system
Scaling LawsScaling LawsPower-law relationships between model performance and parameter count, data volume, and compute
ScratchpadScratchpadAn Agent's working memory space for storing intermediate reasoning steps
Semantic CacheSemantic CacheA technique for caching LLM query results based on semantic similarity (rather than exact matching)
SFTSupervised Fine-TuningModel fine-tuning using labeled data
Short-term MemoryShort-term MemoryConversation history for the current session, limited by the context window
SkillSkillA reusable Agent component encapsulating specific capabilities, including tools, prompts, and execution logic
SSEServer-Sent EventsA protocol for servers to push real-time events to clients
StateStateContext information maintained by an Agent during execution; the core mechanism for data passing between nodes in a Graph Agent
Streamable HTTPStreamable HTTPA new transport protocol introduced by MCP in 2025, supporting on-demand streaming and session resumption
SupervisorSupervisorThe central node in a multi-Agent system that coordinates other Agents
SWE-benchSWE-benchSoftware Engineering benchmark, evaluating an Agent's ability to resolve real GitHub issues
System PromptSystem PromptInitial instructions defining an Agent's behavioral guidelines
TemperatureTemperatureA parameter controlling LLM output randomness (0 = deterministic, 1 = more random)
Text-to-SQLText-to-SQLTechnology for automatically converting natural language descriptions into SQL query statements
TokenTokenThe smallest unit of text processed by an LLM (approximately 1–2 tokens per Chinese character)
ToolToolExternal functionality that an Agent can call (e.g., search, calculation, API calls)
TransformerTransformerThe core architecture of modern LLMs, based on self-attention mechanism, proposed by Google in 2017
uvuvA high-performance Python package manager written in Rust, becoming the new standard for Python package management
Vector DBVector DatabaseA specialized database for storing and retrieving vector embeddings (e.g., ChromaDB, Pinecone)
Working MemoryWorking MemoryAn Agent's temporary reasoning space during task execution, similar to a human's "scratch paper"
Zero-shotZero-shot LearningHaving the model complete tasks through instructions alone, without providing examples