What is an AI agent? A plain-language guide
An agent is more than a chat window. Here is what one is made of, where it works, what keeps its spending in check and how it differs from a chatbot.
Short answer
An AI agent is an AI model with a standing job: instructions that say what to do, skills and connectors that let it reach your tools and information, places where it works and a budget it may spend. A chatbot answers the question in front of it; an agent keeps its job between conversations and can start work on a schedule.
What an AI agent is
What is an AI agent? Ask ten people and you will hear ten answers, from science fiction to a slightly smarter search box. A practical definition is more useful: an AI agent is an AI model that has been given a job, the means to do it and limits to stay within.
The model is the part that reads, reasons and writes. On its own, it only responds to what you type. The agent wraps that model in everything a job needs: a description of the work, access to the information and tools involved, a way to be reached or started, and a budget. Set it up once and it keeps doing the same job, today and next month, without being briefed again.
A useful comparison is a new assistant on their first day. They are capable, but they still need to know what their job is, which systems they may use, where requests come in and what they are allowed to spend. An agent needs exactly the same things, written down once.
The six parts of an agent
Every agent on Atroxabot is assembled from the same six parts. Most of them are choices you make once, when you create the agent.
- Instructions: the job description, in plain language. Who the agent works for, what it should do, what it should never do and how its answers should sound.
- Model: the AI that does the thinking. Atroxabot runs agents on Claude Haiku 5.5, Claude Sonnet 5.5 and Claude Opus 5.5, from fast and light to most capable. Compare the models to see what each is best at.
- Skills: reusable know-how, such as a way of writing a report or a procedure to follow. You write a skill once and assign it to any agent that needs it.
- Connectors: the links between an agent and the services where your information lives, so it can look things up or act there on your behalf.
- Where it works: in a chat in the Atroxabot app, in your messaging apps, or on a schedule in the background. An API and webhooks, for calling agents from your own software, are on their way.
- Budget: the share of your account's credits the agent may use. It can never spend more than that, however busy it gets.
Examples of agents a person might run
Agents are easiest to understand by the jobs they do. The examples below describe the shape of a job, not a particular app: the connectors you give an agent decide which of your services it can reach.
Five agents and the setup that fits each
| Agent | What it does | Where it works | Model that fits |
|---|---|---|---|
| Morning briefing | Gathers what you follow and writes a short summary before your day starts | On a schedule, delivered to a messaging app | Claude Haiku 5.5 |
| Inbox triage | Sorts incoming messages, flags what needs you and drafts routine replies | On a schedule, in the background | Claude Haiku 5.5 |
| Research assistant | Looks into a question, compares sources and writes up what it found | In a chat in the Atroxabot app | Claude Sonnet 5.5 |
| Support helper | Answers customer questions from your own documentation | In a messaging app | Claude Sonnet 5.5 |
| Project planner | Breaks a large goal into steps and works through them over many turns | In a chat in the Atroxabot app | Claude Opus 5.5 |
Illustrations of common jobs. Pick the model by how often the agent runs and how hard its work is.
Notice the pattern in the last column. Agents that run often on light work suit a fast, inexpensive model; agents that research, write or plan need more capability, and pay for it per message. Choosing a model for your agent goes through that decision in detail.
How an agent differs from a chatbot
A chatbot is a conversation. You open it, ask something, get an answer and close it. Whatever it knew about your task lives in that one conversation, and nothing happens until you type again.
An agent is a role. Its instructions persist between conversations, so it does not need to be briefed every time. It can be reached where you already are, such as your messaging apps, and it can start on its own: a scheduled run happens whether or not you open anything. Because it works with less supervision, it also comes with a budget, so a busy week never turns into an unexpected bill.
Chatbot and agent side by side
| What differs | Chatbot | Agent |
|---|---|---|
| Its job | Whatever you ask in this conversation | Standing instructions you wrote once |
| When it works | When you type | When you message it, or on a schedule |
| What it works with | What you give it in the conversation | The skills and connectors you assigned |
| What it may spend | Not set per task | A budget you set per agent |
Setting up your first agent
Write the job down
Describe the work in a few short paragraphs, as you would for a new colleague: the goal, the tone, what good output looks like and what the agent should never do.
Pick a model
Start with the lightest model that can do the job well. Claude Haiku 5.5 is the right first choice for frequent, light work.
Assign skills and connectors
Give the agent only the skills and connectors its job needs. Fewer tools make an agent more focused and easier to trust.
Choose where it works
Decide how work reaches it: a chat in the app, a messaging app, or a schedule that starts it in the background.
Set its budget
Give the agent a share of your credits. How Atroxabot credits work explains what a message costs and how budgets keep spending in check.
You can try all of this for free: the Free plan includes 5,000 credits a month on Claude Haiku 5.5. When you want more credits or the larger models, see the plans.
Frequently asked questions
Is an AI agent the same as a chatbot?
No. A chatbot answers what you ask in one conversation. An agent has standing instructions, skills and connectors to your tools, can be reached in your messaging apps or run on a schedule, and spends within a budget you set.
Do I need to know how to code to make an agent?
No. You set up an agent in the Atroxabot app by writing its instructions and choosing its model, skills, connectors and budget. Code only comes in if you want to call agents from your own software, through an API that is on its way.
Can an agent spend more than I allow?
No. Each agent has a share of your account's credits that you choose, and it cannot use more than that share.
Can several agents share the same skill or connector?
Yes. Skills and connectors belong to your account. You set each one up once and assign it to as many agents as need it.
Which model should my first agent use?
Start with Claude Haiku 5.5 for frequent, light work such as briefings and triage. Move to Claude Sonnet 5.5 when the agent researches, writes or uses several tools, and to Claude Opus 5.5 for long, complex tasks.
