---
language: en
title: "AI agent"
description: "An AI agent performs a task by choosing steps and tools based on feedback from its environment, within defined permissions, completion criteria, and limits."
canonical: "https://tropk.ai/en/resources/glossary/ai-agent"
last_modified: "2026-10-12T16:10:47.142Z"
---

# AI agent

Reviewed on 2026-10-07.

## What it means

An AI agent is a system that pursues a task and chooses steps or tools based on the information it receives while carrying it out. In systems based on language models, this may involve querying data, comparing options, and taking action in connected services. Its autonomy depends on how the application is designed: its objective, permissions, supervision, and stop conditions define the scope of its actions.

## Why it matters for marketing

For marketing, the concept helps specify work that involves intermediate decisions, rather than simply requesting an answer. Checking a record, preparing a change, and applying it are different capabilities. When commissioning or evaluating an automation, the team should define the expected outcome, the available data, and which actions require review, as well as verify the results returned.

## Example in practice

Suppose an agency asks an agent to check for possible duplicates in its CRM. After querying records with the same name, it would find CNPJs with different root numbers and look up the addresses to clarify the comparison. Based on this evidence, it would present the records as separate companies. The verification task would not include merging records; the decision would preserve them and explain the comparison.

## How to apply or verify

Define an observable task, the permitted tools, and the execution limit. In a test environment, track which queries the agent chooses, how it responds to insufficient data, and when it requests intervention. Check the results in the source system. A tool call may fail or simply produce a proposal; the history should make it possible to distinguish an attempt, a response, and a confirmed change.

## Important distinctions

A routine with fixed steps can use AI without having the autonomy to choose its path. An agent, by contrast, can adapt its path and still require supervision. Cryptographic signatures help verify the origin of requests, but they do not automatically grant permission to modify data. Likewise, authenticated access does not mean unrestricted permission or prove that the task was completed.

## Sources and references

- [Building effective agents](<https://www.anthropic.com/engineering/building-effective-agents>) — Anthropic
- [Web Bot Auth](<https://developers.cloudflare.com/bots/reference/bot-verification/web-bot-auth/>) — Cloudflare
- [Authorization — MCP specification 2026-07-28](<https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization>) — Model Context Protocol

## Continue a leitura

- [AI assistant](</en/resources/glossary/ai-assistant>)
- [Model Context Protocol](</en/resources/glossary/model-context-protocol>)
- [Agentic browsing](</en/resources/glossary/agentic-browsing>)

## About the author

[Vinicius Pacheco](</en/team/vinicius-pacheco>)

Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content.

---

[All pages in Markdown](<https://tropk.ai/sitemap.md>)