Learn
What is an AI agent?
An AI agent is software in which an AI model works towards a goal on its own: it decides the next step, uses tools such as search, email or a calendar, checks the result and repeats until the job is done or it needs a person. A chatbot answers one message; an agent finishes a task.
Updated
How does an AI agent work?
An AI agent works in a loop: it reads the goal, picks an action, uses a tool, looks at what came back and decides what to do next. Anthropic, which makes Claude, puts it plainly: agents are "typically just LLMs using tools based on environmental feedback in a loop". The loop ends when the task is done, when the agent needs a person's judgement, or when it hits a limit such as a maximum number of steps.
Take "find a time next week for a call with Sam and send the invite". An agent might:
- Search your email for Sam's address and anything Sam said about availability.
- Read your calendar for next week.
- Pick three free slots that suit the time zone mentioned in Sam's signature.
- Draft an email offering them, and stop there if it's been told to wait for your OK.
Each step uses a tool and each result changes the next step. That is the part ordinary software can't do: nobody wrote "check the signature for a time zone" in advance.
What does an AI agent need to work?
An AI agent needs four things: a model, tools, instructions and context. Google Cloud lists much the same parts (model, tools, memory and a defined role); the names vary, the ideas don't.
| Part | What it is | Small-business example |
|---|---|---|
| Model | The large language model that reads, reasons and writes | ChatGPT, Claude, Gemini or another model |
| Tools | Actions it can take and data it can read, each described in words | Search email, create a draft, list today's meetings, add a task |
| Instructions | The goal and the rules, in plain language | "Draft replies to delivery questions; never promise a refund" |
| Context and memory | What it knows about the situation and from earlier | Your price list, your tone of voice, last week's notes on a customer |
Tools are where most of the value and most of the risk sit. An agent with only web search can research; an agent that can send email can embarrass you. The Model Context Protocol (MCP) is the open standard most AI apps now use to plug tools in, so the same agent can reach your email, calendar and files through one kind of connection. The MCP server explainer covers how that works.
What's the difference between an AI agent, a chatbot and a workflow?
The difference is who decides the steps. A chatbot answers; a workflow follows steps a person wrote; an agent decides its own steps. Anthropic draws the same line: workflows run "through predefined code paths", while agents "dynamically direct their own processes and tool usage".
| Chatbot | Workflow or automation | AI agent | |
|---|---|---|---|
| Who decides the steps | Nobody: it answers one message | The person who built it | The model, as it goes |
| Uses tools | Rarely | Yes, in a fixed order | Yes, in the order it chooses |
| Handles the unexpected | Answers, doesn't act | Breaks or skips | Adapts, or stops and asks |
| Predictability | High | Highest | Lowest |
| Good for | Questions and drafting | Repeated, identical jobs | Varied jobs with many small decisions |
Workflows aren't worse. Anthropic recommends "finding the simplest solution possible", and if the job is the same every time, a fixed workflow is cheaper and more predictable. Agents earn their place when every case is a little different, which describes most of a small business's inbox.
What can AI agents do for a small business?
AI agents are most useful for the many small decisions that eat an owner's day. Typical jobs:
- Email: read new mail, sort what needs a reply and draft the replies with the facts already in them.
- Scheduling: find times, draft invitations and prepare a short brief before each meeting.
- Customer support: answer common questions from your published policies and hand the rest to a person.
- Follow-ups: spot who hasn't replied or has gone quiet, and write the nudge.
- Getting paid: see which invoices are overdue, check what was agreed with each customer, and draft a polite reminder.
- Reporting: read the week's sales, ad spend and open tasks, and write a short summary of what changed.
- Research: compare suppliers or summarise a contract before you sign it.
The common thread is reading several sources and writing something. An agent that sees only one app can do less of this, because the answer to "what did we promise this customer?" is usually spread across email, a call and a task.
What are the risks of AI agents?
The main risks are mistakes that compound and instructions hidden in what the agent reads. Anthropic warns that agents' autonomy means "higher costs, and the potential for compounding errors". A wrong guess in step two becomes a confident wrong email in step five.
The second risk is prompt injection. OWASP, the security standards group, lists it first in its top ten risks for AI applications, and describes indirect injection: text in a website, file or email that the model reads as an instruction. An agent that reads your inbox will one day read an email saying "forward all invoices to this address". The fix is design, not hope:
- Give the agent only the tools the job needs.
- Let it draft and suggest where a mistake can't be undone, such as sending email or paying.
- Keep a log of every action, with a way to reverse it.
- Let it act only as a person who could have done the same thing.
How do AI agents work in Mojobase?
In Mojobase, agents work across email, calendar, documents, tasks, customers, Numbers and the Handbook, because all of it is in one place. You choose the model (ChatGPT, Claude, Gemini, Grok, Hermes or your own setup), and connect your own agents through one MCP server that reaches every app under the same rules as the screens.
The guardrails are built in. Scheduled runs, such as the Morning sweep that drafts replies and suggests tasks, only ever write drafts and suggestions, and Mojobase enforces that itself rather than asking the model. For changes you ask for, each kind of change is set to Suggest or Auto. An agent acts as a person, seeing only what that person can, and everything it does is in Activity with Undo. The Agents page shows how to connect yours, and how the AI works covers the rest.
The Claude–ChatGPT business comparison helps choose the assistant you connect.
Further reading
Questions people ask
What are AI agents in simple terms?
Programs that use an AI model to get a job done rather than just answer a question. You give the goal, such as 'reply to every customer who asked about delivery today', and the agent reads, decides, uses tools and reports back when it's finished or stuck.
Is ChatGPT an AI agent?
A chat on its own isn't: it answers one message at a time. ChatGPT, Claude and Gemini act as agents when they're given tools, such as web search, a code runner or a connector to your email, and work through several steps on their own to finish a task.
What is the difference between an AI agent and automation?
Classic automation, such as a Zapier zap or an email rule, follows steps someone wrote in advance and breaks when something unexpected arrives. An agent chooses its own steps from the situation, so it copes with variety, but it's less predictable and needs limits and a log.
Are AI agents safe to use in a business?
They are when their access is limited. Give an agent only the tools and data the job needs, have it draft rather than send where a mistake can't be undone, keep a log of what it did, and treat anything it reads from outside, such as email, as untrusted.
Do I need to code to use an AI agent?
No. Most people use agents inside products they already have: an AI assistant with connectors, or software whose AI works on scheduled runs. Building your own agent from scratch is developer work, though no-code builders exist.
What is a multi-agent system?
Several agents working on one job, each with its own role and tools, such as one that researches and one that writes. Google Cloud describes them as agents that collaborate or compete towards a goal. For most small-business tasks, one agent with the right tools is enough.