The AI Agent Glossary: 60 AI Terms Explained in Plain English (2026)

Every AI product page now reads like a spelling test: agentic, MCP, RAG, context window, tool calling, guardrails, orchestration. The ideas behind the words are mostly simple, but nobody explains them in one place, and that gap is why so many people "use AI" without ever handing it a real job.

This AI glossary fixes that. It covers 60 terms you will meet when reading about AI agents, large language models and automation in 2026, grouped by theme and explained in plain English with a concrete example. Bookmark it, share it with your team, and use it alongside our guides to what agentic AI is and how AI subagents work. Where a term matters for using Jobbit, we say so.

Core concepts: the words everyone uses

• Artificial intelligence (AI): software that performs tasks which normally need human intelligence, such as understanding language, recognising images or making decisions. In 2026 the term mostly refers to systems built on large language models.

• Machine learning (ML): the technique of training software on examples rather than writing explicit rules. A spam filter learns from millions of labelled emails; it is not told every rule by hand.

• Deep learning: machine learning using neural networks with many layers. It is what made modern image recognition, speech and language models possible.

• Generative AI: AI that creates new content, text, images, code, audio or video, from a prompt. ChatGPT, Claude, Gemini and Midjourney are generative AI products.

• Large language model (LLM): a neural network trained on enormous amounts of text that predicts the next token in a sequence and, in doing so, learns to reason, write and follow instructions. GPT, Claude, Gemini, Llama, DeepSeek and Mistral are LLM families.

• Foundation model: a large model trained on broad data that can be adapted to many tasks. LLMs are foundation models for language; there are equivalents for images and video.

• Multimodal model: a model that understands and produces more than one type of input, typically text, images, audio and sometimes video. Multimodal models let an agent read a screenshot or a scanned invoice.

• Prompt: the instruction you give a model. Prompting well is a skill; see how to write prompts for AI agents.

• System prompt: hidden instructions set by the product builder that shape how the model behaves in every conversation: its role, tone, rules and tools.

• Inference: the act of running a trained model to produce an answer. Training happens once; inference happens every time you send a message, and it is what you pay for.

How models think: tokens, context and reasoning

• Token: the unit models read and write, roughly three quarters of an English word. Pricing, limits and speed are all measured in tokens. "One million tokens" is about 750,000 words.

• Context window: the amount of text a model can consider at once, including your conversation, documents and the model's own reply. Bigger windows let an agent hold a whole codebase or a hundred-page contract in mind.

• Context engineering: the craft of deciding what goes into the context window: which documents, examples, tool results and memories, so the model has what it needs and nothing that distracts it. It has largely replaced "prompt engineering" as the important skill in agent design.

• Reasoning model: a model trained to "think" before answering, producing intermediate steps that improve accuracy on maths, code and planning. Reasoning modes cost more tokens but make agents far more reliable.

• Chain of thought: the step-by-step reasoning a model writes out on the way to an answer. Reasoning models do this natively; older models can be prompted to.

• Temperature: a setting controlling randomness. Low temperature gives consistent, predictable output; higher temperature gives more varied, creative output.

• Hallucination: a confident but false statement from a model, such as a made-up citation or a non-existent product feature. Agents reduce hallucination by checking sources and testing results rather than trusting a single answer.

• Grounding: tying a model's answer to real data, documents, search results or tool outputs, so it reports what is actually there instead of what sounds plausible.

• Fine-tuning: further training a model on your own examples so it adopts a style, format or domain knowledge. Useful for narrow tasks; usually unnecessary for agents, which get their knowledge from tools and context.

• Retrieval-augmented generation (RAG): fetching relevant documents at question time and putting them in the context window so the model answers from your data. It is how "chat with your documents" products work.

• Embeddings: numeric representations of text that capture meaning, so similar passages sit close together. Embeddings power semantic search and RAG.

• Vector database: a database optimised for storing and searching embeddings. Pinecone, Weaviate, pgvector and Chroma are common choices.

• Knowledge cutoff: the date after which a model has no training data. Agents get around it by searching the web and reading current sources.

Agents: the vocabulary of autonomy

• AI agent: software that uses a model plus tools to complete tasks autonomously: it plans, acts, observes results and adapts. The subject of most of this glossary.

• Agentic AI: the broader approach of building goal-driven, tool-using AI systems, often with several agents working together. Explained fully in what is agentic AI?.

• AI assistant / copilot: agent-like help that stays under your direct control, suggesting rather than acting. GitHub Copilot in the editor is the classic example; "agent mode" is where a copilot becomes an agent.

• Autonomy: how much an agent does without asking. Useful systems scale autonomy to the reversibility of the action: draft freely, spend only with approval.

• Tool use / function calling: the mechanism by which a model calls software: search the web, run code, query a database, send an email. Tools turn a chatbot into an agent.

• Model Context Protocol (MCP): an open standard, introduced by Anthropic, for connecting models to tools and data sources in a consistent way. Think of it as a universal plug for agents; support spread across the major AI platforms through 2025 and 2026.

• Agent-to-agent protocol (A2A): a standard for agents from different vendors to discover and talk to each other, proposed by Google. Together with MCP it points toward agents collaborating across companies.

• Orchestration: coordinating multiple steps, tools or agents into one workflow: deciding what runs, in what order, and what happens on failure.

• Multi-agent system: several agents with different roles working on one goal, typically a lead agent that plans and delegates plus specialist agents that research, build or check.

• Subagent: an agent spawned by another agent to handle one piece of a job in parallel. Jobbit uses subagents to research many sources or build and test several features at once; see AI subagents explained.

• Planner: the part of an agent that breaks a goal into steps and revises the plan as results come in.

• Memory: what an agent retains beyond one message. Short-term memory is the current context; long-term memory stores preferences, past decisions and facts about you so the agent improves over time.

• Agent loop: the repeating cycle of plan, act, observe, adjust that runs until the goal is met or the agent needs a decision.

• Human in the loop: a design where a person reviews or approves certain steps, such as sending a customer email or making a payment.

• Computer use / browser agent: an agent that operates software by looking at the screen and clicking and typing like a person, which lets it work with tools that have no API. Explained in computer-use AI agents.

• Deep research: an agentic workflow in which the agent reads dozens of sources, cross-checks claims and produces a cited report. See how to run deep research with AI.

• Workflow automation: predefined, rule-based sequences of actions across apps, the territory of Zapier, Make and n8n. Agents add judgement to workflows; we compare the two in Zapier vs Make vs n8n vs AI agents.

• Multipurpose AI agent: one agent that handles many kinds of work, research, writing, building software, design, automation, from a single chat, rather than a separate tool for each. Jobbit is built this way.

Building software with AI

• Vibe coding: building software by describing what you want in natural language and letting an AI agent write, run and fix the code, judging the result by how it behaves rather than by reading every line. Full guide: what is vibe coding?.

• AI app builder: a product that turns a description into a working application, often with hosting included. Lovable, Bolt, v0, Replit and Jobbit are examples; compared in best vibe coding tools.

• Coding agent: an agent that works inside a codebase: reading files, writing code, running tests and fixing failures. Claude Code, OpenAI Codex, Cursor's agent mode and Devin are coding agents.

• No-code / low-code: visual tools for building software without (or with little) programming, such as Bubble, Webflow, Airtable and Retool. How they compare with vibe coding is covered in vibe coding vs no-code vs low-code.

• Sandbox: an isolated environment where an agent can run code safely without touching your real systems. Essential for any agent that executes code.

• Deployment: putting software live on a server so people can use it. Agents that deploy for you, with hosting included, remove the step that stops most non-developers.

• API (application programming interface): the structured way one piece of software talks to another. Agents use APIs to read and change data in other systems.

• Webhook: a message one system sends to another when something happens ("new order placed"), commonly used to trigger automations.

• MVP (minimum viable product): the smallest version of a product that lets real users try it. Agents have collapsed MVP timelines from months to days; see how to build a SaaS with AI.

• Technical debt: shortcuts in code that make future changes slower and riskier. AI-generated code accumulates it quickly if nobody reviews structure, one of the risks covered in vibe coding mistakes and security.

Safety, quality and trust

• Guardrails: rules and checks that limit what an agent may do or say: blocked actions, spending caps, content filters, required approvals.

• Prompt injection: an attack in which text the agent reads (a web page, an email, a document) contains instructions designed to hijack it. Robust agents treat everything they read as data, never as commands.

• Alignment: the broad effort to make models behave in line with human intentions and values, from refusing harmful requests to following instructions faithfully.

• Evaluation (evals): systematic tests of model or agent quality on a fixed set of tasks. Serious teams run evals before trusting an agent with a workflow.

• Benchmark: a public evaluation used to compare models, such as SWE-bench for software engineering tasks. Useful signals, but real-world performance on your task matters more.

• Observability: logging what an agent did, which tools it called and why, so you can audit results and debug failures.

• Red teaming: deliberately attacking a model or agent to find weaknesses before someone else does.

• Data residency and privacy: where your data is stored and who can see it. Check whether an AI provider trains on your inputs, and where its servers are, before uploading anything sensitive.

Business and pricing terms

• Credits: the usage currency many agent platforms sell instead of raw tokens, so one credit might cover a message, a build step or a research run. Read the conversion carefully; we break it down in how much do AI agents cost?.

• Rate limit: the cap on how many requests or tokens you can use in a period. Hitting it mid-task is the most common reason an agent run stalls.

• Latency: how long a response takes. Agents trade latency for thoroughness: a two-minute research run beats a two-second guess.

• Seat-based vs usage-based pricing: paying per user per month versus paying for what you consume. Agent products increasingly mix the two.

• Automation ROI: the return on an automation, usually measured in hours saved, faster response times or revenue influenced, against the cost of the tool and the time to set it up.

• Digital worker / AI employee: marketing language for a level-4 agent that owns a role rather than a task. Treat the phrase with the same scepticism you would apply to a CV.

Frequently asked questions

What is the difference between an AI agent and an LLM?

An LLM is the model that reads and writes text. An AI agent is a system built around a model that adds planning, tools, memory and a feedback loop so it can complete tasks rather than only answer questions. The LLM is the engine; the agent is the car.

What does MCP mean in AI?

MCP stands for Model Context Protocol, an open standard for connecting AI models to tools and data sources in a consistent way. Instead of every product building custom integrations, tools expose an MCP server and any compatible agent can use them.

What is a context window and why does it matter?

The context window is how much text a model can consider at once, measured in tokens. It matters because everything the agent knows about the current job, your instructions, the documents, the tool results, has to fit inside it. Larger windows let agents handle bigger codebases and longer documents.

What is the difference between RAG and fine-tuning?

RAG retrieves relevant documents at question time and gives them to the model, so it answers from current data without retraining. Fine-tuning changes the model itself by training it on examples, which suits fixed styles or formats. For most businesses RAG plus good tools is enough, and agents make both less necessary by simply reading the source.

What is vibe coding in one sentence?

Vibe coding is building software by describing what you want and letting an AI agent write, run and fix the code, so you judge the result by how the app behaves rather than by reading every line.

Now that the vocabulary makes sense, put it to work: start free on Jobbit and hand an agent a job worth doing.