The parts of agentic AI, in plain words
People say ‘AI’ for several different things. Agentic AI is a few parts working together, and most of what separates it from a chatbot sits outside the model. In one line each:
- The model thinks.
- The harness lets it act.
- Tools are the actions it can take, and connectors reach your other services.
- Permissions set what it may do without asking.
- The knowledge network is what it remembers.
Model
The trained system that reads and writes text. Claude, GPT and Gemini are families of models. A model on its own can only answer: it cannot open a file or run a calculation, and it remembers nothing between conversations. Frontier models are the most capable available at a given time, from the few companies that build them; near-frontier means a step behind. The same model can sit behind a chatbot or an agent. What changes is what is built around it.
Chatbot
A model in a chat window. You type, it replies, and you carry anything useful out by hand. Most chat apps now add a few tools, such as web search or reading a file you upload, but the work still happens one message at a time, and each conversation starts largely from scratch.
Context window
Everything the model can see at once: the conversation so far and any files it has been given. It is large but finite, and it empties when the session ends. That is why a chat forgets: the memory was never in the model.
Scaffold, or harness
The software around a model that lets it act. A harness gives the model tools, carries out what it asks for, shows it the results and lets it try again, in a loop, until the task is done or it needs you. Developers call it a code harness because it grew up around writing software; Claude Code and OpenAI's Codex are examples. Agentic AI is a model working inside a harness.
Agent
A model in a harness, working on a task: reading, doing, checking and correcting over many steps, instead of answering in one.
Tools
The actions a harness gives the model: read and write files, search them, run code and commands, look things up on the web. Wide tool access by default means the agent can use these as the work needs, without asking each time, inside the permissions you set.
Connectors
Links from the agent to the services where your work lives: email, calendar, shared drives, databases. Many use the Model Context Protocol (MCP), an open standard, so one connector works with several harnesses. What a connector can reach is set by its own sign-in and permissions.
Permissions
The rules for what the agent may do on its own, what it must ask about first, and what it may not do at all. Working in a largely autonomous capacity means setting these so the agent can get on with a well-defined task, while anything risky or irreversible still comes back to you.
Knowledge network
The notes, conventions, code and descriptions of your data that the agent reads at the start of each session and adds to as it works, kept as plain files in folders you own. It is how an agent remembers: tomorrow starts where today stopped. We call a well-kept one a knowledge tree.
Why the difference matters for analytical work: chatbots and agentic AI.