You can build a small app with AI before you understand every word people use to describe it. Then a tutorial mentions a harness, a hook, or a swarm, and the conversation suddenly sounds more complicated than the work.
Here are 15 terms worth knowing if you use AI regularly and occasionally build things with it. They fit together: what the model is, how an agent works, what it can use, and what happens when several agents work together.
What the model is working with
1. Weights. The numbers an AI model learns during training. They shape how it responds to new input. You don’t need to inspect them to use ChatGPT or Claude, but the term matters when people talk about downloading or modifying a model. Stanford HAI’s Artificial Intelligence Glossary has a more technical definition.
2. Open weights. A model’s trained weights are available to download. That gives people the option to run the model on their own hardware or a server they control. Check the license before assuming you can use or change it however you like. OpenAI open-weight models (gpt-oss) gives one licensing example; How to Run AI on Your Own Computer: The Complete Ollama Guide shows why you might want to try a local model.
3. Tokens. The small pieces of text a language model reads and produces. A token might be a word, part of a word, or punctuation. Long chats and documents use more tokens, which affects context limits and, in many services, cost. Hugging Face’s Tokenizers lesson shows how text becomes pieces a model can process.
4. Multimodal. Able to work with more than one type of input or output, such as text, images, audio, or video. If you upload a screenshot and ask an AI what is wrong with your page, you’re using a multimodal capability. See Stanford HAI’s Artificial Intelligence Glossary.
5. AGI. Short for artificial general intelligence: the idea of an AI with broad, human-level ability across many kinds of work. People disagree about exactly what would qualify, so treat a claim that a product has “reached AGI” as a claim to examine. Stanford HAI explains the disagreement in What is AGI (Artificial General Intelligence)?.
How an agent gets work done
Imagine asking an AI tool to build a simple webpage. These five terms describe what happens after your request.
6. Agent. An AI system that can choose steps and use tools to work toward a task. It might read your project files, edit the page, run a check, and use the result to decide what to do next. If you ask Claude Code to fix a broken link on your site, it reads the file, finds the problem, edits the URL, and verifies the fix — each step is its own decision. Anthropic describes that cycle in Trustworthy agents in practice; I go deeper in The 4 Parts of Every AI Agent.
7. Goal. The result you want the agent to reach. “Make the page work on a phone and check that it loads” gives it a clearer finish line than “improve the site.” A goal with a verifiable condition — “the tests pass,” “the page loads in under two seconds” — lets the agent know when to stop instead of continuing to tinker. Some tools make this explicit: Keep Claude working toward a goal explains how Claude Code checks a completion condition.
8. Loop. The repeated cycle of taking a step, seeing the result, and deciding what to do next. For the webpage, that could be edit → preview → fix → preview again. You can often watch this happen: the agent writes a line of code, runs it, sees an error, and tries a different approach — that’s the loop in action. You may also hear loop used for a recurring task that starts again on a timer; the meaning depends on the tool and conversation. Trustworthy agents in practice covers the first use, and Keep Claude working toward a goal compares recurring /loop runs with goal-based work.
9. Harness. The software that runs the agent’s work cycle around the model. When the model asks to open a file, the harness carries out that action, sends the contents back, and starts the next step. Claude Code and Cursor both use Claude, but they feel different to work with — that difference is largely the harness. The word is used more broadly in some discussions; Harness, Scaffold, and the AI Agent Terms Worth Getting Right maps both uses.
10. Hook. An action set to run when a particular event happens. In Claude Code, a hook can run a check after a tool finishes or react when the agent tries to stop. You could set one to automatically format your code every time the agent saves a file — it fires without you asking. Think of it as a rule attached to a moment in the workflow, rather than another request you must remember to type. Automate actions with hooks has examples.
What an agent can use or create
11. Skills. Reusable instructions for a particular kind of work. A skill could tell an agent how to review a webpage using your checklist, so you don’t have to explain the process each time. If you find yourself pasting the same three paragraphs of instructions into every conversation, that’s a skill waiting to happen. Claude Code’s Glossary defines its skill format. Stop Re-Explaining Yourself to Claude — Build a Skill Instead shows how to make one.
12. MCP. Model Context Protocol, a standard way for AI applications to connect with outside tools and information. An MCP connection could let an agent look up a design in Figma or check an issue tracker while working on your page. Without MCP, you’d copy information out of one tool and paste it into the chat yourself — MCP lets the agent reach it directly. What the agent can reach depends on the connector and its permissions. What is the Model Context Protocol (MCP)? shows everyday examples.
13. Artifacts. In Claude, an Artifact is a self-contained result—such as a document, design, or small interactive tool—that opens beside the chat for editing and reuse. Ask Claude to build you a quiz or a simple calculator and it appears as an artifact you can use right there. In ChatGPT, the closest current equivalents are writing blocks for editable text and code blocks for code and supported previews. You may also see older ChatGPT tutorials call its editing workspace Canvas. These interfaces overlap in purpose, but their features differ. See What are artifacts and how do I use them?, Working with writing blocks and code blocks in ChatGPT, and the older Introducing canvas page.
When more than one agent is involved
14. Subagent. A smaller worker given part of a larger task. A main agent might ask one subagent to inspect the page’s layout and another to check its links, then use their findings. You can think of it like delegating: the main agent stays focused on the overall job while subagents handle the parts in parallel. In Claude Code, each subagent works in its own context and reports back to the main conversation. Claude Code’s Glossary explains that setup.
15. Agent swarm. Several agents collaborating and handing work among themselves. A customer-service swarm might route your request to a billing agent, which calls a refund agent, which updates your account — each one specialised. Google Cloud uses swarm for a specific design where specialized agents can communicate with one another and an explicit condition tells the work when to end. The phrase also gets used more loosely for multi-agent setups. Choose a design pattern for your agentic AI system explains the technical pattern.
The idea can sound unsettling when agents are connected to the open internet. The practical questions are what they can read, what they can change, and what needs a person’s approval. A swarm with access to your accounts deserves careful boundaries; its name alone doesn’t tell you what those boundaries are. Trustworthy agents in practice describes tool permissions and human control.
When the next unfamiliar term appears, place it on this map. Then ask what the tool can access, what action it can take, and how you’ll know the work is done.