Agentic AI Glossary

Agent - An AI system that can work toward a goal rather than simply answer a question. An agent can reason about what needs to be done, decide which actions to take, use tools, observe the results, and continue until the task is complete. For example, an agent could research a company, retrieve financial data, analyze it, and produce a report.

Agentic AI - AI designed to perform tasks with a degree of autonomy. Instead of only generating an answer to a prompt, an agentic system can plan a sequence of steps, make decisions, use tools, interact with external systems, and adapt based on results. Agents are one way of implementing agentic AI.

AI Agent Loop - The repeating cycle an AI agent uses to accomplish a task. A typical loop is: understand the goal → decide what to do → use a tool or take an action → observe the result → decide what to do next. The loop can repeat many times until the goal is achieved or the agent determines that it cannot proceed.

Alignment - The effort to make an AI system behave according to its intended goals, instructions, and constraints. Alignment can involve training, system instructions, safety mechanisms, human feedback, and other techniques. In practical applications, alignment often means making sure the AI reliably does what the developer and user actually intend.

API (Application Programming Interface) - A defined interface that allows one software system to communicate with another. An API specifies things such as what requests can be made, what information must be provided, and what response will be returned. AI applications commonly use APIs to access databases, financial services, search engines, email systems, and other software.

Artificial Intelligence (AI) - The broad field of creating computer systems capable of performing tasks associated with human intelligence. These tasks can include understanding language, recognizing images, making predictions, reasoning, planning, and generating content. Machine learning and generative AI are important parts of modern AI.

Benchmark - A standardized test used to evaluate and compare AI models. Benchmarks can measure things such as mathematical ability, coding, reasoning, language understanding, or knowledge. A benchmark score can be useful, but it does not necessarily predict how well a model will perform for a particular real-world application.

Chain-of-Thought - A term describing step-by-step reasoning used to solve a problem. Complex tasks can often be solved more effectively when the model reasons through intermediate steps rather than jumping directly to an answer. Modern AI systems may use internal reasoning without exposing their complete reasoning process to the user.

Connector - An integration that connects an AI system to an external application, service, or source of information. A connector might allow an AI application to access Gmail, Google Drive, Slack, a CRM, or a database. Conceptually, a connector provides the bridge between the AI and an external system.

Computer Vision - The field of AI concerned with understanding images and video. Computer vision systems can identify objects, read text, recognize patterns, analyze scenes, and interpret visual information. Modern multimodal models can combine computer vision with language understanding.

Context Window - The amount of information an AI model can process as context during a single interaction. This can include the user's prompt, previous conversation, documents, tool results, and other information supplied to the model. A larger context window allows a model to work with more information at once, although more context does not automatically mean better results.

Deep Learning - A branch of machine learning based on neural networks with many layers. Deep learning became particularly powerful because it can automatically learn complex patterns from large amounts of data. Modern image recognition, speech recognition, and LLMs rely heavily on deep-learning techniques.

Embedding - A numerical representation of information that captures aspects of its meaning or characteristics. For example, a sentence can be converted into an embedding represented by a long list of numbers. Similar pieces of information tend to have embeddings that are mathematically close to one another, making embeddings useful for semantic search and recommendation systems.

Few-Shot Learning - Giving an AI model a small number of examples in the prompt to demonstrate how a task should be performed. The examples help establish the desired format, behavior, or reasoning pattern without retraining the model. For example, you can show three examples of how customer emails should be classified and then ask the model to classify a fourth.

Fine-Tuning - Additional training performed on an already-trained AI model to make it better suited to a particular task, domain, or style. Instead of building a model from scratch, developers start with an existing model and train it further using specialized examples. Fine-tuning can change the model's behavior more deeply than simply changing the prompt.

Foundation Model - A large, general-purpose AI model trained on broad datasets that can serve as the starting point for many applications. LLMs such as GPT-style models are examples of foundation models. Developers can build applications on top of them using prompting, RAG, tool calling, fine-tuning, and other techniques.

Function Calling - A mechanism that allows an AI model to request execution of a predefined software function. The model typically produces structured arguments, while the application actually executes the function and returns the result to the model. For example, an AI could request get_stock_price("AAPL"), with the application responsible for actually retrieving the price.

Generative AI - AI that can create new content rather than simply classify or analyze existing information. It can generate text, code, images, audio, video, and other forms of content. Chatbots based on LLMs are one of the most familiar examples of generative AI.

Guardrails - Rules, checks, and technical controls designed to keep an AI system within acceptable boundaries. Guardrails can validate model outputs, restrict which tools an agent may use, prevent certain actions, or require human approval. They are particularly important when AI can take real-world actions.

Hallucination - A situation where an AI generates information that appears convincing but is incorrect, fabricated, or unsupported by available evidence. Hallucinations can happen because language models generate likely outputs rather than inherently verifying every statement against reality. Techniques such as RAG, tool calling, validation, and structured outputs can help reduce them.

Inference - The process of running a trained AI model to produce an output. For an LLM, inference happens when you send a prompt and the model generates a response. Training creates the model; inference is using the model.

Inference Cost - The computational cost of running an AI model to generate an output. It depends on factors such as model size, number of tokens processed, hardware, and how efficiently the model is served. For applications with millions of users, inference cost can be a major part of the overall operating cost.

Latency - The amount of time it takes for an AI system to respond to a request. Latency can include the time needed to send the request, run the model, call external tools, and return the response. Agentic systems can have higher latency because one user request may involve many model and tool calls.

LLM (Large Language Model) - An AI model trained on very large amounts of text and other data to understand and generate language. LLMs can perform tasks such as answering questions, summarizing documents, writing code, translating languages, and reasoning about problems. An LLM by itself generally generates information; an agent combines an LLM with additional capabilities such as tools, memory, and control logic.

Machine Learning (ML) - A branch of AI in which systems learn patterns from data rather than being explicitly programmed with every rule needed to perform a task. For example, instead of programming every rule for identifying spam, a machine-learning system can learn patterns from examples of spam and legitimate messages.

MCP (Model Context Protocol) - An open protocol designed to standardize how AI applications connect to external tools and data sources. MCP provides conventions for exposing capabilities such as tools, resources, and prompts to AI applications. It can reduce the need to build a custom integration for every AI client and external service.

MCP Server - A server or program that exposes capabilities to an AI application using the Model Context Protocol. An MCP server might expose tools for querying a database, accessing files, interacting with a SaaS application, or executing business operations. The AI client can discover the available capabilities and invoke them according to the MCP protocol.

Multimodal AI - AI that can understand or generate multiple types of information, or "modalities," such as text, images, audio, and video. For example, a multimodal model might receive a photograph and a written question and answer the question about the photograph. This allows AI applications to work with information beyond plain text.

Neural Network - A machine-learning model made from interconnected mathematical units commonly called neurons. The network learns by adjusting numerical parameters so that its outputs become more accurate. Neural networks are the underlying technology behind many modern AI systems, including deep-learning models and LLMs.

Open-Source Model - An AI model whose components are released under licenses that provide some level of access, modification, or redistribution. The exact meaning of "open source" varies in AI because a model may release its weights without releasing its complete training data, training code, or other components. Therefore, it is useful to check exactly what a particular model makes available.

Orchestration - The process of coordinating multiple AI models, agents, tools, services, and workflow steps to accomplish a larger task. For example, an application might use one model to classify a request, another to generate an answer, and several tools to retrieve information. Orchestration determines how these components work together.

Parameter - A numerical value learned by an AI model during training. Parameters influence how the model transforms its input into an output. The number of parameters is one measure of model size, although a larger parameter count does not automatically mean a better model.

Pretraining - The initial large-scale training process used to create a foundation model. The model is exposed to enormous amounts of data and learns general patterns in language, images, code, or other information. This produces a general-purpose model that can later be adapted for specific applications.

Prompt - The instructions, questions, data, or other input provided to an AI model. A prompt can be as simple as a question or as complex as a detailed set of instructions containing examples, constraints, and reference information. The quality and structure of the prompt can significantly affect the model's output.

Prompt Engineering - The practice of designing prompts to make an AI model produce more useful, reliable, or consistent results. It can involve clear instructions, examples, output formats, constraints, and relevant context. Prompt engineering is especially useful when building applications around general-purpose models.

RAG (Retrieval-Augmented Generation) - A technique in which an AI system retrieves relevant information from an external knowledge source and provides that information to the model before generating an answer. This allows an application to use information that may not have been part of the model's training data. RAG is commonly used for company documents, knowledge bases, manuals, and frequently changing information.

Reinforcement Learning (RL) - A machine-learning approach in which a system learns by taking actions and receiving rewards or penalties. The system gradually learns which actions tend to produce better outcomes. Reinforcement learning is widely used in areas such as robotics, games, and some stages of AI model training.

RLHF (Reinforcement Learning from Human Feedback) - A technique for improving AI models using feedback from humans. People evaluate or rank model outputs, and that feedback is used to train the system to produce responses that humans generally prefer. RLHF has been an important technique in the development of conversational AI.

Semantic Search - Search based on the meaning of information rather than just matching exact words. For example, a semantic search for "how do I get my money back?" could find a document titled "Refund Policy" even though the words do not exactly match. Semantic search commonly uses embeddings to compare the meaning of queries and documents.

Skill - A reusable set of instructions, knowledge, and sometimes workflows that teaches an AI system how to perform a particular type of task. A skill is generally higher-level than a tool: a tool performs an individual capability, while a skill can describe how to combine capabilities to accomplish a broader task. For example, a "financial analysis" skill might instruct an agent how to retrieve financial data, analyze it, and produce a report.

Speech Recognition - Technology that converts spoken language into text. It allows applications to understand voice commands, transcribe meetings, create captions, and process spoken conversations. Modern speech-recognition systems can handle multiple languages, accents, and conversational speech.

Temperature - A setting that influences how an AI model selects its next output tokens. Higher values generally allow more variation and creativity, while lower values tend to produce more predictable output. The exact behavior depends on the model and the available generation settings.

Text-to-Speech (TTS) - Technology that converts written text into spoken audio. TTS systems can generate natural-sounding voices and are commonly used in voice assistants, accessibility software, audiobooks, and conversational AI.

Token - A unit of text that an AI model processes. A token might represent a complete word, part of a word, punctuation, or another piece of text. Token counts are important because context-window limits and many AI API costs are measured in tokens.

Tokenization - The process of converting text into the tokens that an AI model can process. For example, a sentence is broken into a sequence of tokens before being passed to an LLM. Different models can use different tokenization schemes.

Tool - A capability that an AI system can invoke to perform an action or obtain information outside the model itself. Examples include searching the web, querying a database, sending an email, executing code, or placing a trade. The AI decides when a tool is useful, but the actual operation is normally performed by external software.

Tool Calling - The ability of an AI model to request that an external tool or function be executed. The model typically identifies the tool and supplies structured arguments, the application executes it, and the result is returned to the model. Tool calling is a fundamental building block for AI agents.

Training - The process of teaching an AI model by exposing it to data and adjusting its parameters. During training, the model repeatedly makes predictions, measures errors, and changes its parameters to improve. Training a large foundation model can require enormous amounts of computing power and data.

Transformer - A neural-network architecture that uses mechanisms called attention to process relationships between pieces of information. Transformers are the architecture behind most modern LLMs and have also become important in vision, audio, and multimodal AI. Their ability to efficiently process relationships between tokens was a major breakthrough for modern AI.

Vector - An ordered list of numbers used to represent something mathematically. In AI, vectors are often used to represent the meaning or characteristics of text, images, users, products, and other objects. Mathematical operations can then be used to compare these representations.

Vector Database - A database optimized for storing and searching vectors. It can quickly find vectors that are mathematically similar to a query vector. Vector databases are commonly used to implement semantic search and the retrieval component of RAG systems.

Weights - The learned numerical values inside a neural network that determine how strongly different inputs influence its behavior. A model's weights contain much of what the model has learned during training. When people refer to downloading or running an "open model," they are often referring specifically to obtaining its model weights.

Zero-Shot Learning - Asking an AI model to perform a task without providing examples of that task in the prompt. For example, you could simply tell an LLM, "Classify this email as spam or not spam," without showing it any previous classifications. Modern foundation models are capable of many zero-shot tasks because they learned broad capabilities during pretraining.