What each thing is
AI model
An AI model is the learned mathematical component. It receives input and produces output, such as a prediction, generated text, or a requested tool call. The model does not, by itself, provide a chat screen, retain application state, or execute an external action. Those are system responsibilities around the model.
AI chatbot
An AI chatbot is an application whose main interface is conversation. It accepts messages and returns conversational responses. Its backend might be one model call, a fixed workflow, an agent, or a mixture of those. “Chatbot” tells you how a user interacts with the product, not how autonomous its backend is.
AI harness
An AI harness is the software and configuration that makes a model usable inside an application. It can build prompts, load relevant context, expose tools, execute tool requests, preserve state, stream output, enforce permissions, and decide when a run must stop.
The term is still settling. Some developers use “harness” specifically for the model-and-tool loop. Others include nearly all of the surrounding application infrastructure. In both uses, the harness is outside the learned model.
AI agent
An AI agent is a system that pursues a goal by letting a model choose actions, observe results, and choose what to do next. The harness runs and constrains that process. An agent often uses tools and may work across several steps without asking the user to direct every one.
There is no universal autonomy threshold. A useful practical test is control flow: if the model can adapt the next step to an intermediate result, the system is agentic. If code fixes the complete path in advance, it is better described as a workflow, even when the workflow makes several model calls.
The actual distinction
Each term answers a different question:
- Model: What learned component turns an input into an output?
- Chatbot: Where does a person hold a conversation with the system?
- Harness: What software supplies context, executes operations, and applies controls around the model?
- Agent: What system chooses a sequence of actions in pursuit of a goal?
| Term | Kind of thing | Main responsibility | Can it affect an external system by itself? |
|---|---|---|---|
| Model | Learned component | Produce an output from an input | No. It can request or describe an action, but software must execute it. |
| Chatbot | Application and interface | Conduct a conversation | It depends on its backend. A simple chatbot only responds; an agentic chatbot can act. |
| Harness | Software control and execution layer | Prepare model calls, run allowed tools, manage state, and enforce rules | It performs configured operations, but does not supply the model’s judgment or goal by itself. |
| Agent | Goal-directed system | Select and perform steps based on observations | Yes, within the tools, permissions, and approval rules its harness provides. |
The categories can nest. A chatbot can use a harness to call a model. If the model can direct a tool loop toward a goal, that same product also contains an agent. An agent can instead run from an API request, a schedule, or an event without any chatbot at all.
One request, four roles
Suppose you type: “Move my delivery to Friday and confirm the change.”
The chatbot receives the message and shows the response. The model interprets the request and might decide that it needs to look up an order. The harness constructs the model input, offers an order lookup tool, validates the requested call, runs it, and returns the result.
If Friday is unavailable, an agent can adjust. It might inspect other dates, ask you to choose one, wait for your reply, update the delivery after approval, and then verify the new date. The model chooses those next steps from the observations. The harness is what actually calls the delivery system and applies the permission rules.
A non-agentic chatbot could still handle the request. It might follow a hard-coded exchange form or simply explain how you can change the date yourself. Conversation alone does not determine agency.
When each distinction matters
When the model matters
Focus on the model when you are comparing the quality and shape of generated outputs: which inputs it accepts, how reliably it follows instructions, or how it behaves on your evaluations. Changing a model can improve or alter decisions without changing the product’s interface or execution design.
When the chatbot matters
Focus on the chatbot when the problem is conversational usability. Message history, turn-taking, clarification, tone, accessibility, and handoff to a person are chatbot concerns. A capable backend can still make a poor chatbot if the interaction is confusing.
When the harness matters
Focus on the harness when the problem involves reliability, context, tools, state, security, or observability. A model cannot enforce its own permissions merely by being instructed to do so. The harness must restrict available operations, validate calls, require approvals, record results, and stop unsafe or runaway execution.
The same model can behave very differently under two harnesses because it sees different context, receives different tools, and gets different feedback after acting.
When the agent matters
Focus on the agent when the task requires choosing and revising a path rather than producing one response. Agency is useful for open-ended work whose intermediate results determine the next step. It also introduces more ways to fail: a mistaken early action can change what happens later, so tool limits, verification, and human approval matter.
What people commonly confuse
A chatbot is not the same as its model
The chatbot is the product surface and application behavior. The model is one component behind it. One chatbot can route between models, and one model can power many unrelated products.
A chat window does not prove there is an agent
A single model response can appear in a chat bubble. Conversely, an agent can run in the background with no conversation. Chat describes the interface; agency describes how the system pursues work.
Tool access alone does not settle the agent question
A fixed program can always call the same tool after every request. That is automation, but the model is not directing the process. The stronger signal is whether model output determines which action happens next and whether observations can change the plan.
The harness is not the intelligence inside the model
The harness may contain substantial logic, but it does not become part of the trained model. It decides what the model can see and do, executes requests, and applies deterministic controls. Keeping that boundary visible makes failures easier to diagnose: a poor decision may be a model problem, while a bad tool result or missing approval check is a system problem.
“Agentic chatbot” is not a contradiction
A product can be both. It is a chatbot because conversation is its interface. It is agentic when its backend can choose and execute multiple steps toward the user’s goal.
How the pieces fit together
A common arrangement looks like this:
user → chatbot surface → harness → model
↕
tools and dataWhen the model uses tool results to select further actions toward a goal, the model, harness, and available tools together form an agent. The chatbot remains an optional surface on top.
Read the component pages for the full treatment of AI models, AI chatbots, AI harnesses, and AI agents. Continue to What Is Tool Use in LLMs? for the boundary between a model proposing an action and the harness executing it.