Skip to main content

Glossary

Terms and definitions used throughout this catalog, organized alphabetically.

A​

Agent​

A system that uses a language model to perceive its environment, make decisions, and take actions toward a goal. Agents typically have access to tools and maintain some form of memory. See Agents.

Agentic System​

An AI application exhibiting autonomous behavior — goal-directed actions, tool use, planning, and self-correction with minimal human intervention. Exists on a spectrum from simple tool use to open-ended goal pursuit.

Autonomy​

The degree to which an agent operates without human intervention. Ranges from human-approval-at-every-step to fully autonomous execution. Higher autonomy increases throughput but reduces control.

C​

Chain of Thought (CoT)​

A reasoning technique where the model shows step-by-step logic before reaching a conclusion. Improves accuracy on complex tasks by making intermediate reasoning explicit. See Chain of Thought.

Citation​

Linking generated output back to source documents or data. Provides verifiability and reduces hallucination risk. See Citation.

Context Rot​

The gradual degradation of model performance as the context window fills with irrelevant or stale information. A key symptom of the Amnesiac Agent anti-pattern.

Context Window​

The maximum number of tokens a language model can process in a single request — system prompt, conversation history, tool results, and response combined. Larger windows don't guarantee better attention across all content.

D​

Delegation​

Passing a subtask from one agent to another. Central to Orchestrator and Hierarchical patterns.

E​

Embedding​

A numerical vector representation of text that captures semantic meaning. Used for similarity search in Vector Store memory and RAG.

Eval Suite​

A collection of test cases, scoring rubrics, and benchmarks used to systematically measure agent output quality. See Eval Suite.

Evaluator-Optimizer​

An orchestration pattern where a generator produces output and a critic scores it in a refinement loop. See Evaluator-Optimizer.

F​

Fan-Out / Fan-In​

Splitting a task into concurrent subtasks (fan-out) and aggregating results (fan-in). The core mechanism of the Parallel pattern.

Few-Shot Prompting​

Including examples in the prompt to demonstrate desired behavior. Contrast with zero-shot (no examples) and many-shot (large numbers of examples).

Function Calling​

The ability of a language model to output structured requests to execute external functions. The mechanism behind Tool Router and all tool integration patterns.

G​

Grounding​

Connecting model outputs to external sources of truth — databases, search results, documents — to improve factual accuracy. RAG and Citation are grounding patterns.

Guardrails​

Validation layers that check agent output for safety, compliance, factual accuracy, or format correctness before delivery. See Guardrails.

H​

Hallucination​

When a model generates false or unsupported information with apparent confidence. Mitigated by RAG, Citation, and Guardrails.

Human-in-the-Loop​

Inserting human approval or review checkpoints into an agent workflow. Trades throughput for control at high-stakes decision points. See Human-in-the-Loop.

J​

JSON Schema​

A vocabulary for defining the structure and constraints of JSON data. Used by Structured Output to enforce typed responses from language models.

K​

Knowledge Graph​

A graph-structured representation of entities and their relationships. Used as a memory pattern for complex, interconnected information. See Knowledge Graph.

L​

LLM (Large Language Model)​

A neural network trained on large text corpora that can generate, analyze, and transform text. The reasoning engine at the core of agentic systems.

LLM-as-Judge​

Using a language model to evaluate and score the output of another model or agent. See LLM-as-Judge.

M​

MCP (Model Context Protocol)​

A standardized protocol for dynamic tool discovery and invocation. Allows agents to find and use tools at runtime without hardcoded integrations. See MCP.

Memory​

State that persists across agent interactions. Categories include conversation buffers, summaries, vector stores, extracted facts, file-based persistence, knowledge graphs, and shared memory. See Memory patterns.

Multi-Agent System​

An architecture using multiple specialized agents that coordinate to accomplish tasks beyond any single agent's capability. Includes Role-Based, Hierarchical, and Swarm structures.

O​

Observation​

In the ReAct loop, the result returned by a tool after an action is executed. Feeds back into the agent's next reasoning step.

Orchestration​

The coordination of multiple agents or processing steps — routing requests, managing workflows, aggregating results, and handling failures. See Orchestration patterns.

P​

Pipeline​

A fixed sequence of processing steps where each step's output feeds the next, with optional validation gates between stages. See Pipeline.

Plan-and-Execute​

A reasoning pattern where the agent creates an explicit plan before taking action, then executes steps sequentially with optional replanning. See Plan-and-Execute.

Prompt​

Instructions given to a language model. Includes system prompts (persistent instructions), user prompts (requests), and assistant messages (model outputs).

Prompt Injection​

An attack where untrusted input manipulates agent behavior by embedding instructions that override the system prompt. A key concern in the Happy Path Mirage anti-pattern.

R​

RAG (Retrieval-Augmented Generation)​

A pattern that retrieves relevant documents or data before generation, providing context the model can reference. Reduces hallucination and enables access to current or private information. See RAG.

ReAct​

A reasoning pattern that interleaves Thought, Action, and Observation steps in a loop. The agent reasons about what to do, executes a tool, observes the result, and repeats. See ReAct.

Reflection​

A reasoning pattern where the agent critiques its own output and revises it. Can be self-reflection (same model) or cross-reflection (separate critic). See Reflection.

Router​

An orchestration pattern that classifies incoming requests and dispatches them to specialized handlers. Single-pass decision with no iteration. See Router.

S​

Sandbox​

An isolated execution environment for running agent-generated code safely. Prevents file system access, network calls, or other side effects. Central to Code Execution.

Scoring Rubric​

A structured set of criteria used to evaluate agent output. Defines what "good" looks like across dimensions like accuracy, completeness, and safety. Used in LLM-as-Judge and Eval Suite.

Streaming​

Delivering agent output progressively as tokens are generated rather than waiting for the complete response. See Streaming.

Structured Output​

Constraining model output to conform to a predefined schema — JSON, XML, or typed objects. See Structured Output.

Summarization​

Compressing conversation history or documents into shorter representations that preserve key information. See Conversation Summarization.

Swarm​

A multi-agent architecture where peers self-organize without a central controller. Agents communicate through shared state and handoff protocols. See Swarm.

System Prompt​

Persistent instructions that define an agent's role, constraints, and behavior. Remains constant across user interactions within a session.

T​

Temperature​

A parameter controlling randomness in model output. Lower values (0.0–0.3) produce deterministic, focused responses. Higher values (0.7–1.0) increase creativity and variation.

Token​

The basic unit of text processing for LLMs. Roughly 3–4 characters in English. Pricing, rate limits, and context windows are all measured in tokens.

Tool​

An external capability an agent can invoke — APIs, databases, code execution, file systems, or other agents. Tools extend what agents can accomplish beyond text generation.

Tracing​

Recording the full sequence of an agent's reasoning steps, tool calls, and decisions. Essential for debugging, evaluation, and observability in production systems.

V​

Validation Gate​

A checkpoint in a Pipeline that verifies output quality, format, or safety before passing data to the next stage. Fails fast on bad input.

Vector Database​

A database optimized for storing and querying embeddings by similarity. Powers semantic search in RAG and Vector Store memory.

W​

Worker​

A specialized agent that handles a specific subtask assigned by an Orchestrator. Workers are typically narrow in scope and stateless.

Workflow​

A defined sequence of steps or agent interactions to accomplish a task. Can be fixed (pipeline) or dynamic (orchestrator-driven).

Z​

Zero-Shot​

Prompting a model to perform a task without providing examples. Relies entirely on the model's pre-trained knowledge and instruction-following ability.