What is an AI agent?

A chat assistant answers your question. An agent does a job for you. Say "collect last month's invoices and draft reminder emails for the unpaid ones", and an agent connects to the accounting system, pulls the invoices, separates the unpaid ones, writes the emails and presents them for your approval. If a step fails, it tries another way.

At the center of an agent sit a language model and a loop:

• Plan: the model breaks the goal into steps.
• Act: it calls a tool, such as an API, a database, a web browser or an email system, using function calling.
• Observe: it reads what the tool returned.
• Decide: it checks whether the goal is reached and picks the next step.

The loop runs until the job is done or the agent needs human approval, and the agent keeps track of what it has done in its memory.

Common examples today are assistants that handle customer requests and book appointments or route them to the right team, systems that classify incoming email and open CRM records, coding agents that write code and run tests, and research agents that search the web and prepare reports. Coordinating several agents that work together is called orchestration.

The AI agents QANATONE builds are of this kind: they answer incoming demand around the clock, qualify it and route it to the right person.

Because an agent acts in real systems, its mistakes have real consequences, such as an email sent to the wrong person or an invoice issued for the wrong amount. Agents therefore get only the permissions they need, critical steps like payments and customer messages require human approval, and every action is logged. All of this is part of the guardrail layer. Agents can also still lose their way on long, open ended tasks.