Concept guide · Understand agents
An AI agent does more than answer. It works through a goal using tools.
A chatbot waits for the next message. An agent can plan several steps, use approved tools, inspect what happened and continue until it reaches a stopping point.
A normal example
Imagine a supplier-comparison agent. You give it thirty supplier websites and the fields you care about. It visits each source, collects the same information, marks missing details, creates a spreadsheet and stops for your review.
The AI is not “an employee.” It is software following a goal with more freedom than a fixed script.
Chat, automation or agent?
You want one answer, draft or explanation and can stay in the conversation.
The steps and rules are stable: when a form arrives, copy fields and send a fixed confirmation.
The five parts to design
- A narrow goal with a visible finish line.
- The information it may use.
- The tools it may call, such as search, email or a database.
- The actions that require your approval.
- A stop condition for success, failure, cost or uncertainty.
The human bit is in the design
You choose the goal, permissions, quality checks and point where a person takes over. “Review the result” is not enough. Name what must be checked: five sampled sources, every price, any customer-facing message and all actions that change another system.
If a spreadsheet formula, form rule or scheduled script can do the job reliably, use that. An agent adds variable behaviour, model cost and more ways to fail.
Try it without building anything
- Finish this sentence: “The agent is done when…”
- List the information it can read.
- List the actions it can take.
- Circle every action that needs approval.
- Write one reason it must stop and ask for help.