# What is an AI agent?

> An AI agent pursues a goal by chaining steps and tools on its own. Definition, the agent loop, a worked example for a Quebec SME and the guardrails it needs.

Updated on October 15, 2026 · https://app.pilotiko.com/en/glossary/ai-agent

An AI agent is an artificial intelligence system that pursues a goal by choosing and chaining its own steps: it consults sources, uses tools, checks the result of each action and adjusts the next one. Unlike a simple text generator, it can act on other systems, which is why limits and human oversight matter so much.

## In short
- Agents work in a loop: act, observe, adjust.
- Most are built on a large language model that calls tools.
- Autonomy must be bounded: limited access, logs and a way to stop.
- Sensitive actions stay subject to prior human validation.

## How do Anthropic and the Quebec government describe agents?

Anthropic separates workflows, where language models and tools follow predefined code paths, from agents, where the model dynamically directs its own process and tool use. In practice, it says, agents are typically [large language models](https://app.pilotiko.com/en/glossary/llm) using tools based on environmental feedback in a loop [1].

Quebec's Ministère de la Cybersécurité et du Numérique published a 2026 guide on agentic AI for public bodies. It describes agentic AI as systems that can pursue an objective, orchestrate steps, consult sources or use tools within a defined scope, and it calls a concrete instance of such a system an AI agent. It stresses that an agent's autonomy is a supervised capability, not general permission to act without human oversight [2].

## What does an agent actually do, step by step?

Anthropic notes that agents need ground truth from the environment at each step, such as the result of a tool call, to assess progress. They can pause for human feedback at checkpoints, and they are usually given stopping conditions such as a maximum number of iterations [1].

1. **Take the goal** — For example: build tomorrow's schedule for the Laval crew.
2. **Look things up** — Read the calendar, pending requests and each technician's workload.
3. **Propose** — Draft assignments and flag conflicts.
4. **Check** — Test the draft against constraints and fix any overlapping appointments.
5. **Hand back to a person** — Present the proposal; nothing is confirmed to a customer without approval.

## How is an agent different from a chatbot or a workflow?

| Tool | What it does | Who decides the steps? |
|---|---|---|
| Chatbot (conversational agent) | Replies to a user in natural language | The conversation, one reply at a time |
| Automated workflow | Runs a sequence of steps known in advance | Whoever programmed the rules |
| AI agent | Pursues a goal by choosing steps and tools | The model, within the limits set |

Anthropic recommends finding the simplest solution that works and only adding complexity when needed, which may mean not building an agent at all [1].

## How much prep time can an agent take on?

**Example: a window-cleaning company with 6 crews in Laval**

Assumption: every evening, the dispatcher builds the next day's schedule by hand. An agent drafts a proposal that she reviews and corrects.

- Appointments to place per evening : 30
- Manual preparation time : 75 min
- Time to review the agent's proposal : 20 min
- Difference per evening : 55 min
- Over 5 evenings a week : 275 min, about 4.6 h
- **= About 4.6 hours a week shifted from building to reviewing, while the final call stays with the dispatcher.**

## Which mistakes should you avoid with AI agents?

- Granting every permission just in case: Quebec's guide says access, data, tools and actions should be limited to what is strictly needed [2].
- Keeping no record: without logs you cannot tell what the agent consulted, proposed or carried out [2].
- Forgetting the off switch: a way to stop, suspend and return to manual mode should be designed in from day one [2].
- Taking the output on faith: like any model, an agent can produce an [AI hallucination](https://app.pilotiko.com/en/glossary/ai-hallucination).
- Replacing a simple [automated process](https://app.pilotiko.com/en/glossary/process-automation) that already worked with an agent.

## What do AI agents do in Pilotiko?

In Pilotiko's virtual office, each department is a room, and AI agents have desks there alongside staff. They read the business's real data and prepare work: assignments proposed from calendars and workload, internal reports, editorial calendar drafts. Any action touching a customer, money or a send waits for [a person's approval](https://app.pilotiko.com/en/glossary/human-in-the-loop).

The phone-based [AI receptionist](https://app.pilotiko.com/en/glossary/ai-receptionist) is not available yet; it is planned.

## Frequently asked questions

### Can an AI agent work alone all night?

Technically it can chain steps without anyone stepping in. In practice it should run within limits: stopping conditions, restricted access and human validation before sensitive actions. Letting an agent act unsupervised on customers or money is not advisable.

### What is the difference between an AI agent and chatting with an AI?

A chatbot answers your messages one at a time. An agent receives a goal and picks the steps and tools to reach it. That means it can act on your systems, not just reply.

### Does an AI agent replace an employee?

It takes over specific tasks, mostly preparation and checking. Judgment, the customer relationship and the final decision remain human. It is more accurate to think of it as added capacity than as a replacement.

### How do I know whether a task suits an agent?

Ask three questions: does the task vary too much for fixed rules, can the result be checked, and could a mistake be undone? If any answer is no, a simpler automation or a person is a better fit.

### What does the OQLF call an AI agent?

Its terminology database defines an agent intelligent as an agent that uses AI resources and learning mechanisms and adapts its actions based on experience [3]. On its French pages, Pilotiko uses the plainer agent IA.

## Sources

1. Anthropic, « Building effective agents », https://www.anthropic.com/engineering/building-effective-agents (accessed on 2026-10-15)
2. Ministère de la Cybersécurité et du Numérique (Gouvernement du Québec), « Guide sur l'utilisation responsable de l'intelligence artificielle agentique », https://cdn-contenu.quebec.ca/cdn-contenu/adm/min/cybersecurite_numerique/Intelligence_artificielle/GU_IA_Agentique.pdf (accessed on 2026-10-15)
3. Office québécois de la langue française (Vitrine linguistique), « agent intelligent (GDT) », https://vitrinelinguistique.oqlf.gouv.qc.ca/fiche-gdt/fiche/8394387/agent-intelligent (accessed on 2026-10-15)