Blog · Agents
What is an AI agent, and how is it different from a chatbot?
An AI agent is a language model that has been given tools and is allowed to use them in a loop until a task is done. A chatbot answers your message and stops; an agent reads your goal, decides on a step, takes it, looks at the result and decides what to do next.
That one difference, acting rather than only replying, changes what the software is good for, what can go wrong and how much supervision it needs. This guide explains the parts of an agent, where the line with a chatbot really sits, and what to check before you let one work on anything that matters.
The short version
| Chatbot | Agent | |
|---|---|---|
| What it produces | Text for you to read | Text plus actions in other systems |
| How many steps | One reply per message | Many steps per request |
| Who carries out the work | You, by copying the answer somewhere | The agent, through its tools |
| What it can touch | Nothing outside the conversation | Whatever its tools allow |
| Main risk | A wrong answer | A wrong action |
Both are usually built on the same kind of model. The model is not what makes something an agent. The surrounding program is.
The parts of an agent
Most agents have the same four parts.
A model. A large language model reads text and produces text. On its own it cannot open a file, send an email or run a command. It can only write.
Tools. A tool is a function the surrounding program offers to the model, described in words: "read this file", "search the web", "run this shell command", "create a calendar event". When the model wants to use a tool, it writes a structured request. The program, not the model, carries out the request and passes the result back.
A loop. The program sends the model the goal and the list of tools. The model replies with either a final answer or a tool request. If it is a tool request, the program runs it, adds the result to the conversation and asks the model again. This repeats until the model says it has finished, a limit is reached or a person stops it.
Context. Everything the model knows about the task sits in its context: your instructions, the files it has read, the results of earlier steps. Context is limited, so agents summarise, take notes or read files again as they go. Some keep memory between sessions in ordinary files.
Put together, the pattern is simple: think, act, observe, repeat. That is all "agentic" means in practice.
A worked example
Ask a chatbot, "Why is the test suite failing?" and paste in an error. It will read the error and suggest likely causes. You then go and check each one yourself.
Ask a coding agent the same question inside your project. It can run the tests, read the failing output, open the file the error points to, read the function, form a theory, make a change, run the tests again and report what it found. If its first theory was wrong, the next test run tells it so and it tries something else.
The chatbot gave advice. The agent did the investigation. The agent's answer is also grounded in what it actually saw, because it looked at the real files rather than guessing from a pasted fragment.
Where the line blurs
The boundary is not sharp. Many chat products can now search the web or run a short piece of code before they answer, which is a small agent loop inside a chat window. It is more useful to think of a scale:
- Plain chat. Text in, text out.
- Chat with a few tools. The model can look something up or do a calculation, then replies.
- Task agents. The model works through many steps on your files, inbox or systems, with a person checking in.
- Long-running agents. The model works for an extended period on a larger goal, often alongside other agents.
The further along the scale, the more the questions shift from "is the answer good?" to "what is it allowed to do, and who checks?"
What agents are good at, and where they fail
Agents suit work that has clear feedback. Code is the best example, because tests pass or fail. Research with sources that can be checked, tidying structured data, and preparing drafts for a person to review all work well for the same reason: there is something to check the result against.
They are weaker where success is hard to verify, where a small misunderstanding early on compounds over many steps, or where an action cannot be undone. Agents make the same kinds of mistake a chatbot makes. They can misread an instruction or state something false with confidence. The difference is that an agent's mistake can become a deleted file or a sent message rather than a sentence on a screen.
Agents can also be misled by what they read. A web page or document can contain text written to look like instructions. A well-built agent treats what it reads as information rather than orders, but this is a known weak point and a good reason to limit what an agent can do without review.
Questions to ask before you use one
- What tools does it have? An agent is exactly as powerful as its tools. Read the list. See skills, plugins and tool servers explained.
- Which actions need a person? Reading is usually safe. Sending, deleting, paying and publishing usually deserve a confirmation. See designing approvals people do not skip.
- Can its work be undone? Working on a copy, keeping version history and preparing drafts rather than sending are the usual answers.
- Whose account does it run under, and what leaves your machine? See what actually gets sent to an AI provider.
- Is there a record? You should be able to see afterwards what the agent did and who approved it.
Where Prism fits
Prism Desktop is a desktop workspace where people and AI agents work together in the same channels, threads and direct messages. It runs the AI vendors' own agent programs side by side, each signed in with the owner's own account, and each given only the skills, plugins and tool servers you choose. Agents propose; a person approves a merge, an email or an install, and the decision is recorded. It is Linux first, with early access by waitlist: you can join the waitlist.