| Agent | AI that can take steps on its own, not just answer | Researching flights and filling in a booking form for you to approve |
| API | A behind-the-scenes connection that lets software talk to an AI | How an app adds a “summarize” button powered by someone else’s AI |
| Automation | A “when this happens, do that” rule that runs a task on its own. AI can make it smarter, like drafting a reply | A new sign-up gets added to your spreadsheet automatically |
| Benchmark | A standard test used to measure and compare what AI models can do | A set of coding problems every new model is scored on |
| Containment (sandbox) | A walled-off test space meant to keep an AI from reaching anything outside it | Letting an agent practice on a pretend computer instead of real systems |
| Context | The background information you give the AI | “I’m a florist, and my customers are mostly brides.” |
| Context window | How much the AI can keep in mind at once in one chat | Why a very long chat may “forget” what you said at the start |
| Frontier model | The most advanced AI models available at any given time | The newest top models from companies like Anthropic, OpenAI, and Google |
| Generative AI | AI that creates new things: text, images, audio, video | Writing a blog draft or making an illustration |
| Grader | The program or person that scores an AI’s work during training or testing | A checker that marks an answer right or wrong |
| Guardrails | Safety limits built into an AI to keep it from doing harmful things | Refusing to help someone break into a computer system |
| Instance | One running copy of a model. Many instances of the same model can work at once | A thousand agents, all the same model, each on its own task |
| LLM (large language model) | The kind of AI behind chat assistants, trained on huge amounts of text | The technology inside Claude, ChatGPT, and Gemini |
| Memory | A feature that lets an assistant remember facts about you between chats | Remembering that you run a bakery |
| Multi-agent system | Many AI agents working together and coordinating, sometimes called a swarm | Agents splitting up a big research project and sharing notes |
| Multimodal | AI that handles more than text, like images, audio, or files. Not to be confused with “multi-model,” which means a tool that uses several different AI models behind the scenes. | Uploading a photo of a receipt and asking what you spent |
| Open model (open-source model) | AI that anyone can download and run themselves | Often used by developers or privacy-focused businesses |
| Reinforcement learning | Training an AI by rewarding results you want, a bit like praising a pet for a trick | Scoring answers so the AI learns which kind to give |
| Reward hacking | When an AI finds a shortcut that earns the reward without doing the task the way people intended | Changing a test so it passes, instead of fixing the problem |
| Training | How an AI learns, by studying enormous numbers of examples | Reading billions of pages before you ever use it |
| Training data | The examples an AI learned from | Books, websites, articles, code |
| Token | A chunk of text, about three-quarters of a word, that AI counts | Why some tools have limits like “100,000 tokens” |
| Web search | A feature that lets an assistant look things up online before answering | Checking today’s business hours instead of guessing |
