Quick answer: A chatbot answers questions from what it already knows and hands the reply back to you. An AI agent uses tools, takes multiple steps, and works toward a goal, doing things instead of just describing them. Every agent can chat, but most chatbots cannot act.
The line between the two blurs in marketing copy, where every chatbot is suddenly an "AI agent." So let me draw the line clearly, then show you the difference with a concrete example, then tell you when a plain chatbot is actually the smarter choice.
The difference in one sentence
A chatbot is a conversation. An agent is a conversation that can reach out and change the world.
A chatbot takes your message, runs it through a language model (sometimes with a knowledge base attached), and returns text. The loop is one step long: you ask, it answers. It cannot check anything, cannot call an API, cannot do the thing it is describing.
An agent takes a goal, decides what actions would move toward it, executes those actions with tools, looks at the results, and decides what to do next. The loop runs until the goal is met. The model is the same underneath; the wrapper around it is completely different.
Side by side
| Dimension | Chatbot | AI Agent |
|---|---|---|
| Job | Answer and converse | Accomplish a goal |
| Steps | One turn per reply | Many steps per task |
| Tools | Usually none | Search, fetch, APIs, code |
| Knowledge source | Training data plus a fixed knowledge base | Whatever it can retrieve live |
| Can it act? | No, it only produces text | Yes, it calls tools that do things |
| Handles "then do X"? | No | Yes |
| Predictability | High | Lower, needs guardrails |
| Cost per interaction | Low | Higher, multiple model calls |
Before and after: a concrete example
Imagine a user asks a support system: "Is my order still going to arrive Friday given the storm warnings in Denver?"
A chatbot answers from what it knows. It has your order date and a shipping estimate in its knowledge base. It cannot see the weather, cannot see the carrier's live tracking, so it guesses from the static estimate:
"Your order is scheduled to arrive Friday based on standard shipping times."
That answer sounds confident and might be wrong. The chatbot has no way to check.
An agent treats the same question as a goal and reaches for tools. Roughly, it does this:
Goal: determine if order #4821 arrives Friday given Denver weather
Step 1 call tracking_api(order="4821")
-> package in Denver facility, "weather delay possible"
Step 2 call web_search("Denver weather Friday shipping delays")
-> storm system, regional carrier delays reported
Step 3 reason over both results
Answer "Your package reached the Denver facility but the carrier has
flagged a weather delay, and storms are forecast through Friday.
Saturday is more realistic. I can notify you when it moves."
Same question, completely different behavior. The chatbot recited an estimate. The agent checked reality, combined two live sources, and gave an answer grounded in facts that did not exist when either model was trained.
The capability that separates them
Notice what made the agent useful in that example: it could pull live data. That is the deciding capability. An agent that cannot reach the outside world is a chatbot with extra steps.
The two most common tools an agent needs are web search and web fetch, because so many goals reduce to "find out what is true right now" and "read this specific page." The web is where current facts live.
This is exactly the plumbing link.sc provides. One call fetches any URL and returns clean markdown, and one call searches the web and returns full page content instead of snippets:
curl https://link.sc/v1/fetch \
-H "Authorization: Bearer lsc_..." \
-d '{"url": "https://carrier.example.com/track/4821", "format": "markdown"}'
Wire that behind a tool your agent can call, and the "before" chatbot becomes the "after" agent. For the step-by-step version, see give your AI agent internet access.
Why the word "goal" matters
The cleanest test I know: does the system have a goal, or just a question to answer?
A chatbot is question-shaped. Every reply stands alone. A good chatbot for "what's your refund policy" never needs a second step.
An agent is goal-shaped. It holds an objective across several actions and keeps working until it is done or gives up. "Process this refund" is a goal: check eligibility, verify the order, issue the credit, send confirmation. Each step depends on the last. No single reply finishes it.
If you can imagine the task done in one message, you want a chatbot. If getting it done requires a sequence of dependent actions, you want an agent. For more on the actions themselves, read what is tool calling in LLMs.
When a chatbot is the better choice
Do not build an agent because agents are fashionable. For a large share of real use cases, a chatbot is faster, cheaper, and more reliable, and reaching for an agent just adds points of failure.
A chatbot is the right pick when:
- The answers live in a fixed knowledge base (policies, docs, FAQs).
- Each interaction is self-contained and does not depend on the last.
- You need predictable, auditable responses with low latency.
- There is nothing external to check and nothing to do.
A support bot that answers "how do I reset my password" from your help docs should stay a chatbot. Wrapping it in a tool-calling loop makes it slower and gives it more ways to go wrong for zero benefit.
Reach for an agent when the task needs live data, external actions, or several dependent steps. Handling a return, researching a topic across sources, monitoring a system and reacting: those are goals, not questions.
The honest summary
A chatbot talks. An agent talks and acts. The difference is not the model, it is whether the system can take steps, use tools, and pursue a goal past a single reply.
Start with a chatbot. Upgrade to an agent only when a real task cannot be finished in one turn, and when you do, give the agent live web access, because most goals worth automating depend on facts the model was never trained on.
Turning your chatbot into an agent that can actually check things? link.sc is the fetch and search tool that gives it live eyes on the web. Start free with 500 credits a month.