A reference of key terms used throughout this wiki and in the broader agentic AI ecosystem.
Agent
A system where an LLM operates with a degree of autonomy, dynamically directing its own processes and tool usage to accomplish open-ended tasks. Agents decide on their own what steps to take, rather than following a fixed sequence.
Augmented LLM
The foundational building block of agentic systems: an LLM enhanced with retrieval, tool use, and memory capabilities. Most agentic patterns are compositions of augmented LLMs rather than bare language models.
Chain of Thought
A prompting technique where the model is encouraged to show its reasoning step by step before arriving at a final answer. Chain of thought improves accuracy on complex tasks by making the model’s reasoning process explicit and inspectable.
Context Window
The maximum amount of text (measured in tokens) that a model can process in a single request. The context window includes the system prompt, conversation history, tool results, and the model’s own output. Managing context window usage is a key engineering concern in agentic systems.
Embedding
A numerical vector representation of text that captures its semantic meaning. Embeddings enable similarity search, clustering, and retrieval-augmented generation by allowing systems to find text that is conceptually related, not just lexically matching.
Function Calling
The mechanism by which an LLM requests the execution of a specific function with structured arguments. The model outputs a function name and parameters in a defined format, and the host application executes the function and returns the result. Also referred to as tool calling.
Grounding
The process of connecting an LLM’s outputs to verifiable external sources of truth, such as databases, documents, or API results. Grounding reduces hallucination by anchoring the model’s responses in factual data rather than relying solely on its parametric knowledge.
Hallucination
When a model generates information that is plausible-sounding but factually incorrect or fabricated. Hallucination is a fundamental challenge in LLM-based systems and a primary motivation for techniques like grounding, RAG, and human-in-the-loop review.
Human-in-the-Loop
A design pattern where a human reviews, approves, or corrects an agent’s actions at critical decision points. This is especially important for high-stakes or irreversible operations where the cost of an error is significant.
LLM (Large Language Model)
A neural network trained on large volumes of text data that can generate, analyze, and transform text. LLMs serve as the reasoning engine at the core of agentic systems.
Multi-Agent System
An architecture where multiple agents collaborate, each with distinct roles, capabilities, or areas of expertise. Agents may communicate through shared state, message passing, or an orchestrating coordinator.
Orchestration
The process of coordinating multiple components, steps, or agents to accomplish a complex task. Orchestration can be handled by deterministic code (workflows) or by an LLM that dynamically decides how to delegate and sequence work.
Prompt Engineering
The practice of designing and refining the instructions, examples, and context given to an LLM to elicit desired behavior. In agentic systems, prompt engineering extends to system prompts, tool descriptions, and the framing of intermediate reasoning steps.
RAG (Retrieval-Augmented Generation)
A pattern where relevant information is retrieved from an external knowledge source and included in the model’s prompt before generation. RAG grounds the model’s output in specific, up-to-date data and is one of the most common ways to reduce hallucination.
Reasoning
The cognitive process by which an LLM analyzes information, draws inferences, and reaches conclusions. Reasoning quality is influenced by prompting technique, model capability, and the structure of the information provided in context.
Reflection
A pattern where an LLM reviews and critiques its own output, identifying errors, gaps, or areas for improvement. Reflection is a core component of evaluator-optimizer loops and self-correcting agent architectures.
Tool Use
The ability of an agent to invoke external functions, APIs, or services to perform actions beyond text generation. Tool use is what transforms an LLM from a conversational model into an agent capable of affecting the external world.
Workflow
An agentic system where the sequence of LLM calls and tool invocations is orchestrated through predefined code paths rather than by the model itself. Workflows trade flexibility for predictability and are appropriate when the task structure is known in advance.