# AI agent

URL: https://ayonix.com/glossary/ai-agent
Updated: 2026-09-08 | Published: 2026-09-08 | Author: Jan Mocary, Chief Technology Officer, Ayonix AI

An AI agent is a program that takes a goal, plans the steps to reach it, calls tools or APIs to carry them out, and checks its own progress before reporting back. Unlike a chatbot, an agent acts on systems rather than only producing text about them.

## What are the parts of an AI agent?

The tool list is the security boundary. An agent can only affect what it has been given a tool for, so the review question in an enterprise deployment is which tools exist and what each one is allowed to touch, rather than how capable the underlying model is.

- A model that decides what to do next, given the goal and everything observed so far.
- Tools: the functions, APIs and databases the agent is permitted to call, each with a defined signature.
- Memory: what carries between steps within a task, and what persists between tasks.
- A control loop that decides when to continue, when to ask a human, and when to stop.
- Logging of every step, without which no one can reconstruct why the agent did what it did.

## How is an agent different from a chatbot?

Chatbot and agent compared on what each one does with a request.

The failure mode is why the two are governed differently. A chatbot that answers wrongly costs a reader some time; an agent that acts wrongly has changed a record, sent a message or moved money, and the control that matters is which actions require confirmation before they happen.

- Output — Chatbot: Text — Agent: Actions, then text
- Steps per request — Chatbot: One — Agent: Many, decided at runtime
- Touches other systems — Chatbot: No — Agent: Yes, through defined tools
- Failure mode — Chatbot: A wrong answer — Agent: A wrong action, possibly irreversible
- What review needs — Chatbot: Answer quality — Agent: An audit trail of steps and calls

## What does an enterprise deployment have to control?

- Scope each tool to the narrowest permission that still allows the task, using a service identity rather than a person's credentials.
- Require confirmation before any irreversible action, and define exactly which ones qualify.
- Log every step, tool call and result in a form an auditor can read months later.
- Set a budget in steps or time, so a loop terminates rather than running until someone notices.
- Decide where the model runs, since sending internal data to a hosted model is a data transfer whatever the interface looks like.

## Frequently asked questions

**What is an AI agent?**
An AI agent is a program that takes a goal, plans the steps to reach it, calls tools or APIs to carry them out, and checks its progress before reporting back. The distinguishing property against a chatbot is that it acts on systems rather than only producing text about them.

**What is the difference between an AI agent and a chatbot?**
A chatbot produces text in a single step. An agent decides a sequence of steps at runtime and calls tools that change things: sending a message, updating a record, running a query. That difference in effect is why agents need action-level permissions and an audit trail, and chatbots generally do not.

**What tools can an AI agent use?**
Only the ones it is given. Tools are defined functions with a signature and a permission scope: a database query, an HTTP call, a file operation, a message send. The tool list is the security boundary of the deployment, because an agent cannot affect anything it has no tool for.

**Are AI agents safe to run in production?**
They are as safe as the permissions and confirmations around them. The controls that matter are least-privilege tool scoping, a confirmation step before irreversible actions, complete logging of steps and calls, and a step or time budget so a loop terminates on its own.

**Can an AI agent run on-premise?**
Yes, when the model, the tools and the orchestration all run on local hardware. Ayonix ATLAS agents are built for that case, including air-gapped installations. The practical constraint is model size against available hardware, not the agent architecture itself.

**Do AI agents replace staff?**
In deployments that work, they take the mechanical parts of a workflow and leave the judgement and the approval with a person. Handing an agent an unsupervised end-to-end process is what produces the failures that get written about, because there is no step at which a mistake is caught.