AI Glossary

Agent memory is the information an AI agent can keep or retrieve across tasks, sessions, or workflows so it can understand context and continue work.

An agent multiplexer is a tool that lets you run, monitor, and coordinate multiple AI agents from one interface.

Agent orchestration is the coordination of multiple AI agents working together, managing which agent runs when, what information it receives, and how results are passed between agents.

An agent registry is a place where AI agents, tools, skills, connectors, or resources can be listed, described, verified, and discovered.

Agent runtime isolation means separating the environments where AI agents run so one agent cannot access, modify, or interfere with another agent's code, data, tools, or sessions.

An agent skill is a reusable instruction set that helps an AI agent perform a repeated task consistently, with defined steps, rules, and output format.

Agent washing is when a product is marketed as an AI agent when it's actually just a basic chatbot, a simple automation, or a standard workflow tool with minimal AI autonomy.

An agent-friendly website is designed so AI agents can easily read, navigate, and interact with it, finding information, completing tasks, and taking actions without confusion.

Agentic AI refers to AI systems designed to act autonomously toward goals, planning, using tools, and taking sequences of actions rather than just responding to single prompts.

Agentic engineering is the practice of using AI coding agents as tools while the human engineer stays in control, reviewing outputs, verifying behavior, and remaining accountable.

Agentic Resource Discovery is an emerging open specification for publishing, discovering, and verifying AI tools, agents, skills, and connectors across registries.

An agentic workflow is a process where one or more AI agents autonomously carry out a series of steps to complete a task, making decisions along the way rather than waiting for a human at each step.

An AI agent is a system that can take a goal, break it into steps, and complete those steps autonomously using AI and connected tools.

AI agent infrastructure is the technical layer that lets AI agents work beyond simple chat, including tools, connectors, memory, scheduled tasks, permissions, logs, registries, and discovery systems.

AI alignment is the challenge of ensuring that AI systems pursue goals and produce behaviors that match human intentions and values, not just technically, but in practice.

An AI assistant is a software application that uses artificial intelligence to understand natural language, answer questions, and help users complete tasks through conversation.

AI automation uses artificial intelligence to perform repetitive or complex tasks automatically, reducing the need for manual human effort. It goes beyond traditional automation by handling unstructured data and making judgment calls.

AI code risk review is a way of reviewing AI-generated code based on what the code can break, not just how many lines changed.

An AI coding agent is an AI system that can write, edit, test, and debug code autonomously, going beyond simple code suggestions to actually completing coding tasks end-to-end.

An AI connector links an AI assistant to an outside app, file source, or service, such as Gmail, Google Drive, Slack, Notion, a CRM, or a calendar.

An AI copilot is an AI assistant integrated directly into a tool or application to help users complete tasks within that tool, like GitHub Copilot for code or Microsoft Copilot in Office.

AI ethics is the study and practice of ensuring AI is developed and used in ways that are fair, transparent, accountable, and beneficial, minimizing harm and respecting human rights.

AI evals (evaluations) are tests designed to measure how well an AI model performs on specific tasks, helping developers, researchers, and users compare models and spot weaknesses.

AI export controls are government rules that restrict who can access certain AI models, chips, software, or capabilities based on country, nationality, or security concerns.

An AI gateway is a middleware layer that sits between your application and multiple AI model providers. It manages routing, rate limiting, cost control, logging, and fallbacks, so you don't have to build that logic into every application separately.

An AI goal defines what success looks like for an AI task, the outcome, constraints, verification method, and stopping condition.

AI integration means connecting AI capabilities into existing software, workflows, or business processes, adding AI features to tools you already use rather than replacing them.

AI literacy is the ability to understand, evaluate, and effectively use AI tools, knowing what AI can and can't do, how it works at a basic level, and how to use it responsibly.

An AI loop is a repeated process where an AI keeps trying, checking, and improving until a condition is met or it needs human help.

AI model access risk is the business risk that an AI model becomes unavailable, restricted, slower, more expensive, or limited because of policy, pricing, regulation, outages, or provider decisions.

AI model dependency risk is the risk of relying too heavily on one AI model or provider, so your workflow breaks if access, pricing, quality, or availability changes.

AI orchestration is the process of coordinating multiple AI models, tools, and steps into a unified workflow or pipeline to complete complex tasks.

AI Overviews is Google's feature that shows an AI-generated summary answer at the top of search results, above the regular links.

AI permission hygiene is the habit of giving AI tools only the access they truly need, reviewing those permissions regularly, and requiring human approval before risky actions.

AI red teaming is the practice of deliberately trying to make an AI model fail, produce harmful outputs, or behave in unintended ways, in order to find and fix its weaknesses before they cause real-world harm. It's a security and safety testing practice borrowed from cybersecurity.

AI safety is the field of research focused on ensuring that AI systems behave as intended, remain reliable, and don't cause unintended harm, now and as AI becomes more capable.

AI search is a type of search engine that uses AI to generate a direct answer to your question instead of just showing a list of links.

An AI skill is a reusable instruction set, behavior, or capability that helps an AI perform a specific task consistently.

AI vendor lock-in happens when your prompts, workflows, data, agents, or automations become so tied to one AI provider that switching becomes difficult.

An AI video generator is a tool that creates video content from text descriptions, images, or other inputs using AI, without traditional filming or editing.

AI washing is when a company overstates, misrepresents, or uses AI as a convenient explanation for decisions that may actually be driven by other factors.

An AI workflow is a repeatable, step-by-step process that uses AI tools to complete a task more efficiently than doing it manually.

Answer Engine Optimization (AEO) is the practice of structuring content to appear as answers in AI-powered tools like Perplexity, voice assistants, and featured snippets.

An API (Application Programming Interface) is a way for software systems to communicate with each other. AI APIs let developers access AI capabilities from companies like OpenAI and Anthropic in their own applications.

Artificial intelligence is the field of computer science focused on building systems that can perform tasks that normally require human intelligence, like understanding language, recognizing images, and making decisions.

The attention mechanism is the core technique inside transformer AI models that lets them focus on the most relevant parts of an input when generating a response. It's what allows AI to understand context and relationships across long texts rather than processing words one at a time in isolation.

An autonomous agent is an AI system that can independently plan and execute multi-step tasks, making decisions and taking actions without requiring human input at each step.

A benchmark is a standardized test used to measure and compare the performance of AI models. Benchmarks score models on tasks like reasoning, math, coding, and language understanding so people can compare different models objectively.

AI bias refers to systematic errors in AI outputs that favor or disadvantage certain groups, often because the training data reflected existing human biases.

A blind-spot audit is a final AI review step where the model points out weak assumptions, missing context, uncertainty, and what you should verify before acting.

A browser agent is an AI that can control a web browser, clicking links, filling out forms, reading pages, and taking actions on websites, to complete tasks on your behalf.

Chain-of-thought prompting asks an AI to show its reasoning step by step before giving a final answer, which significantly improves accuracy on complex or multi-step problems.

A chatbot is a software program that simulates conversation with users, typically to answer questions or handle common tasks. Modern AI chatbots use LLMs; older ones follow scripted rules.

Chunking is the process of splitting large documents into smaller pieces before feeding them into an AI system. Since AI models have a limited context window and retrieval works better on focused sections, chunking breaks content into manageable, searchable units.

CLAUDE.md is a project instruction file used with Claude Code to store rules, context, coding preferences, and repeated corrections for future AI coding sessions.

Computer vision is the field of AI focused on enabling computers to interpret and understand images and video, recognizing objects, faces, scenes, and visual patterns.

Context engineering is the practice of carefully designing what information you give to an AI model, the instructions, examples, background, and data, to consistently get better results.

A context file is a Markdown or text file that stores instructions, preferences, project details, examples, constraints, or rules that you reuse with AI tools.

Context waste is unnecessary, duplicated, outdated, or irrelevant information that makes an AI session slower, more expensive, or less focused.

A context window is the maximum amount of text an AI model can process and remember in a single conversation or session.

The decide-execute-deliver sandwich is a way of seeing software work in three layers, deciding what to build, executing the implementation, and delivering verified software.

Decision mapping is a planning method where AI separates fixed decisions from unknowns, research gaps, prototypes, expert questions, and parallel work tracks.

Deep learning is a type of machine learning that uses multi-layered neural networks to learn complex patterns. It's the technology behind most modern AI breakthroughs.

A diffusion model is the type of AI behind most modern image generators. It learns to generate images by starting with random noise and progressively refining it into a coherent output.

Distillation is a technique where a smaller AI model is trained to mimic the behavior of a larger, more capable model. The result is a compact model that performs nearly as well as the original but runs faster and costs less.

An embedding is a numerical representation of text (or other data) that captures its meaning. Similar words or sentences have embeddings that are close together in mathematical space.

An endpoint is a specific URL address where an API receives requests. In AI, each endpoint handles a different task, one for generating text, another for creating images, another for transcribing audio.

Enterprise-managed auth is authentication and access management controlled by an organization, often through identity providers, admin settings, roles, permissions, and audit controls.

Few-shot prompting is a technique where you give an AI model a few examples of the input and output you want before asking it to complete your actual task.

Fine-tuning is the process of taking a pre-trained AI model and training it further on a smaller, specific dataset to specialize it for a particular task or style.

A foundation model is a large AI model trained on broad data that can be adapted to many different tasks. GPT-5, Claude, and Gemini are all foundation models.

Function calling is a feature that lets AI models trigger specific functions or APIs in your code, enabling them to take real-world actions beyond generating text.

Generative AI refers to AI systems that can create new content, text, images, audio, video, or code, rather than just analyzing or classifying existing content.

Generative Engine Optimization (GEO) is the practice of structuring your content so AI search engines and AI answer tools are more likely to cite it in their generated responses.

GPT (Generative Pre-trained Transformer) is the AI architecture behind ChatGPT. Developed by OpenAI, GPT models are some of the most widely used large language models in the world.

Grounding in AI means connecting a model's outputs to real, verifiable sources of information, reducing hallucinations and making responses more factually reliable.

Guardrails are safety controls built into AI systems that prevent the model from producing harmful, inappropriate, or off-policy outputs.

AI hallucination is when an AI model generates information that sounds confident and plausible but is factually incorrect or completely made up.

Harness engineering is the work of designing the tools, workflows, verifiers, prompts, memory, and guardrails around an AI model so it performs reliably.

Human-in-the-loop (HITL) is an AI design principle where humans review, approve, or correct AI outputs at key decision points, rather than letting AI act fully autonomously.

Indirect prompt injection is when malicious instructions are hidden inside content an AI reads, like an email, website, PDF, or calendar invite, so the AI follows them without the user realizing.

Inference is the process of using a trained AI model to generate outputs, when ChatGPT responds to your message, it's running inference.

Intent-based design tokens are design tokens named by their purpose or decision rather than their raw value, making design systems clearer for both humans and AI agents.

A knowledge base is a structured collection of information, documents, FAQs, policies, and data, that an AI can be connected to so it can answer questions based on your specific content.

Latency is the delay between sending a request to an AI model and receiving its first response. Low latency means fast responses; high latency means noticeable waiting. It's a key factor when choosing an AI model for real-time applications.

An LLM (Large Language Model) is the AI technology behind tools like ChatGPT, Claude, and Gemini. It's trained on large amounts of text to generate and understand language.

An LLM-safe design system is a design system built so that LLMs and coding agents are guided toward approved UI decisions and blocked from inventing off-system styles.

A local LLM is an AI language model that runs directly on your own computer instead of in the cloud, giving you privacy, offline access, and control over your data.

LoRA is a technique for fine-tuning large AI models efficiently. Instead of retraining all of a model's billions of parameters, LoRA adds small, trainable layers on top, dramatically reducing the time, cost, and computing power needed to customize an AI model.

Machine learning is a type of AI where systems learn from data rather than being programmed with explicit rules. The more data they see, the better they get.

An MCP connector is a connection built around the Model Context Protocol that lets an AI app or agent access external tools, data, or context in a structured, consistent way.

Mixture of Experts (MoE) is an AI model architecture where only a fraction of the model's total parameters are activated for any given input. Instead of using the whole model for every request, a routing mechanism selects the most relevant 'experts', making it more efficient than dense models of similar capability.

A model access restriction is a policy, provider decision, government rule, or rollout limit that controls who can use a powerful AI model.

A model card is a short document published alongside an AI model that describes what it was trained on, what it's designed to do, its known limitations, and how it should (and shouldn't) be used. It's the AI equivalent of a product label.

Model Context Protocol (MCP) is an open standard that lets AI models connect to external tools, files, and data sources in a consistent way. It allows AI assistants like Claude to access your files, apps, and services without custom integrations for each one.

A model fallback is a backup AI model or provider that can take over when your main model is unavailable, too expensive, rate-limited, or not good enough for a task.

Model parameters are the internal numerical values that define what an AI model has learned. When you hear '70 billion parameter model,' it refers to 70 billion of these learned numbers that shape how the model thinks and responds.

Model routing is the practice of sending different parts of an AI workflow to different models based on cost, speed, quality, or task type.

A multi-agent system is an AI setup where multiple AI agents work together, each handling a specific part of a larger task. One agent might plan, another research, another write, and another review, all coordinating to get the job done.

Multimodal AI refers to AI systems that can understand and generate multiple types of content, text, images, audio, and video, not just text.

Natural language processing (NLP) is the field of AI focused on enabling computers to understand, interpret, and generate human language.

A neural network is a type of AI model loosely inspired by the human brain, made up of layers of interconnected nodes that learn to recognize patterns in data.

No-code AI refers to tools and platforms that let users build AI-powered applications and automations without writing any code, using visual interfaces and pre-built components.

On-device AI is artificial intelligence that runs directly on a device, like your phone or laptop, rather than sending data to cloud servers to process.

Open weights refers to AI models whose trained parameters are publicly released so anyone can download, run, and modify them, without needing permission or paying the original creator.

Open-source AI refers to AI models whose code and weights are publicly available for anyone to download, use, modify, and build upon, in contrast to closed, proprietary models.

Portable context is reusable information you keep outside one AI app so you can carry your preferences, project details, writing style, and working rules between different AI tools.

Prompt caching is a feature offered by AI API providers that saves the processed version of repeated prompt content so it doesn't have to be recomputed on every request. It reduces latency and cuts costs for applications that send similar long prompts repeatedly.

Prompt engineering is the practice of writing clear, specific instructions that help AI tools produce better, more useful outputs.

Prompt injection is a security attack where malicious instructions are embedded in content an AI processes, attempting to hijack the AI's behavior or bypass its instructions.

Quantization is a technique that reduces the size of an AI model by storing its numerical values with less precision. A quantized model runs faster and uses less memory while keeping most of its original capability.

Query fan-out is when an AI search tool breaks one user question into multiple separate searches or sub-queries to gather more comprehensive information before generating an answer.

RAG (Retrieval-Augmented Generation) is a technique that connects an AI model to external data sources so it can answer questions based on specific documents or databases.

A reasoning doom loop happens when an AI model keeps thinking, checking, or re-planning without making useful progress.

A reasoning model is an AI language model that works through problems step-by-step before giving its final answer, making it better at complex tasks that require logic, math, or careful analysis.

Reinforcement learning is a type of machine learning where an AI learns through trial and error by receiving rewards for good actions. RLHF (Reinforcement Learning from Human Feedback) is how most modern AI assistants are trained to be helpful.

Reranking is a step in AI search and retrieval pipelines where an initial set of results is re-scored and reordered by a more precise model to surface the most relevant content. It improves the quality of what gets passed to an AI model for answering questions.

RLHF is a training technique where humans rate AI outputs to teach the model what good responses look like. It's how most modern AI assistants, including ChatGPT and Claude, learned to be helpful, honest, and safe rather than just predicting text.

A rogue agent is an AI agent or agent workflow that behaves outside its intended permissions, goals, or security boundaries.

A scheduled AI task is an AI task that runs at a planned time or on a recurring interval, instead of only when a user manually prompts it.

Semantic compression is reducing context length while preserving the important meaning an AI model actually needs.

Semantic search is a search method that finds results based on the meaning of a query rather than exact keyword matches. It understands context and intent, so searching for 'car problems' also surfaces results about 'vehicle repair', even if those exact words don't appear.

Shadow mode is a way to test new logic, AI output, or code in the background without affecting real users, so you can compare it against the live system before trusting it.

A Small Language Model (SLM) is a compact AI model with far fewer parameters than large models like GPT-5 or Claude. SLMs trade some capability for speed, lower cost, and the ability to run on devices like laptops and phones.

Speech-to-text AI converts spoken audio into written text. Also called transcription or automatic speech recognition (ASR), it powers voice assistants, meeting transcription, and dictation tools.

A subagent is a smaller AI agent assigned to one specific part of a bigger task, such as researching, coding, testing, reviewing, or summarizing.

Synthetic data is artificially generated data used to train or test AI models, rather than data collected from real-world events or people. AI companies increasingly use synthetic data to overcome limitations in real data availability, quality, and privacy.

A system prompt is a special instruction given to an AI model before the conversation starts, setting its role, behavior, and rules for the entire session.

A task map breaks a project into smaller tasks and labels each one by whether AI should do it, assist with it, or stay out of it.

A task-specific AI agent is an AI system built to handle one particular type of job, like booking travel, writing emails, or processing invoices, rather than being a general-purpose assistant.

Temperature is a setting that controls how random or creative an AI's responses are. Low temperature produces focused, predictable outputs. High temperature produces more varied and creative ones.

Text-to-image AI generates visual images from text descriptions. You describe what you want in words, and the AI creates an image matching your description.

Text-to-speech AI converts written text into spoken audio. Modern AI TTS systems produce voices that are nearly indistinguishable from human speech.

A token is the basic unit of text that AI language models process. Tokens are roughly equivalent to word fragments, about 3–4 characters each on average in English.

Tokenization is the process of breaking text into smaller pieces called tokens before an AI model can process it. A token is roughly 3–4 characters or about 0.75 words in English.

Tool use refers to the ability of AI models to call external tools, like web search, calculators, code execution, or APIs, to complete tasks beyond pure text generation.

Training data is the dataset used to teach an AI model. The quality and diversity of training data heavily determines what the model knows and how well it performs.

The transformer is the neural network architecture that powers almost all modern AI language models, including GPT, Claude, and Gemini. It was introduced in a landmark 2017 paper.

A vector database stores data as numerical embeddings and allows fast similarity searches, finding the content most semantically similar to a query. Essential for RAG and AI search.

A verification gate is a required check that AI-generated work must pass before it is trusted, shipped, published, or used.

A verifier loop is a feedback loop where an AI agent or external system checks whether generated work behaves as expected, often through tests, evals, CI, or validation steps.

Vibe coding is a way of building software by describing what you want in plain language and letting an AI coding agent write the code, without needing to write it yourself.

A worktree is a separate working copy of a code project that lets developers or AI agents work on different tasks without overwriting each other.

Zero-shot prompting means asking an AI to complete a task without giving it any examples first. The model relies entirely on its training to understand what you want.