What you’ll achieve with this tutorial

The ChatGPT Agent has arrived to transform how users interact with AI assistants. Instead of just answering questions, it executes tasks autonomously: searches the web, fills out forms, navigates websites, creates files, and even interacts with external services, all without you needing to click anything. Think of it as an assistant that receives a mission and pursues the result, rather than asking you for instructions at every step.

To use the feature on mobile, you need a ChatGPT account with Plus, Pro, or Team plan (the free tier doesn’t have agent access as of this publication date), and the official OpenAI app installed on your Android or iPhone, updated to the latest version available in the app stores. The tutorial below was built based on official OpenAI information and publicly documented behavior of the feature.

Requirements before you start

Before looking for the button, confirm you have:

• Active ChatGPT account with Plus plan (from US$20/month, with exchange rate variations in Brazil) or higher
• iOS or Android ChatGPT app updated and downloaded from official App Store or Google Play stores
• Stable internet connection, preferably Wi-Fi, because the agent makes several sequential requests
• Consent to allow external actions: the agent asks for your confirmation before executing irreversible actions, but you need to be aware that it will navigate real websites

Users in regions with access restrictions to certain OpenAI services may encounter availability limitations. In Brazil, OpenAI operates normally, but the agent still has blocks on some website categories due to company policy.

Step 1: Open the app and access the correct mode

1. Open the ChatGPT app on your phone
2. On the main chat screen, tap the model selector, which appears just above the text field (or at the top of the screen, depending on app version)
3. Select GPT-4o or the model listed as compatible with the agent in your account; according to OpenAI, agent mode is available on this model for eligible subscribers
4. Check if the tools option appears or if the “agent” or “tasks” icon is visible in the composition bar

If you don’t see these options, update the app. OpenAI does gradual releases (phased rollout), so there may be a gap between announcement and arrival to your profile.

Step 2: Activate agent tools

1. In the text field, tap the plus (+) icon or the tools symbol next to the message field
2. A list of capabilities appears: Web Search, Data Analysis, Code Execution, Computer Actions (Computer Use, available only for higher plans depending on region) and others based on your subscription
3. Enable the tools that make sense for your task; you can leave all active and the model decides which to use, or restrict manually
4. Confirm any permissions requested by the app, if they appear on first activation

Tools like Computer Use (where the agent controls a remote browser visual interface) may have different availability depending on region and plan. Check OpenAI’s plans page to see what’s included in your subscription.

Step 3: Write a clear and specific instruction

This is where most users go wrong. The agent responds much better to concrete objectives than vague questions.

Examples of instructions that work well:

• “Research the three best prepaid cell phone plans available in Brazil today, compare prices and data allowances, and bring me a table”
• “Access the Brazilian IRS website and explain the steps to issue a tax compliance certificate”
• “Find the PDF manual for the TP-Link Archer AX73 router and tell me what QoS settings are available”

Instructions that produce poor results:

• “Help me with internet”
• “Research something for me”
• “I want to buy a notebook”

The difference is context and scope. The more you define what you want, the lower the chance of the agent taking a different path than expected.

Step 4: Monitor execution and confirm sensitive actions

1. After sending the instruction, the agent starts working and you see an action log on screen, showing what it’s doing in real time: which websites it accessed, what it searched for, what data it collected
2. For actions involving data submission, form filling, or anything the agent classifies as “irreversible,” it pauses and asks for your confirmation before proceeding
3. Read carefully what the agent is about to do before confirming; this is the correction moment
4. If the path is wrong, use the cancel or redirect button and write a new, more specific instruction

This confirmation behavior is intentional according to OpenAI: the system was designed to keep the user in control of external actions, especially those affecting third-party services.

Step 5: Receive and use the result

Upon completion, the agent delivers a consolidated response: text, table, generated file, or combination thereof. You can:

• Ask to refine the result with a new command in the same conversation
• Request that it export the content in specific format (Markdown, list, spreadsheet, depending on active tools)
• Share the conversation using the native share button in the app

The session history is saved to your account, also accessible via ChatGPT on the web, if you need to resume on another device.

Practical use cases for everyday life in Brazil

The agent has real utility in everyday situations. Some examples that work well according to documented behavior of the feature:

• Price research: ask to compare prices of a product in Brazilian stores, with links to where it found them
• Long document summary: you attach a PDF and request structured analysis
• Professional email drafts: with detailed context, it writes and you review
• Information monitoring: in sessions with search tools active, it tracks what you asked and organizes the data

For those who use their phone as a work tool and spend considerable time researching information, comparing options, or producing content, the agent delivers real value. The learning curve is in writing good instructions, but once you get the hang of it, the time savings are notable. By the way, if you use your phone together with other connected smart devices, Perplexity’s Comet browser is another AI tool worth attention for integrating search and navigation differently from ChatGPT.

Limits you need to know

The agent is not unlimited and has important restrictions:

• Number of actions per session: OpenAI establishes usage limits that vary by plan; upon reaching the limit, you receive a warning and can continue in a new window or wait for renewal
• Websites with bot blocking: many Brazilian sites (especially banks and government services) block automated access, which can prevent the agent from completing certain tasks
• Real-time data is not always accurate: the agent may bring outdated information from websites that don’t update frequently; always confirm critical data from the official source
• Privacy: actions performed by the agent go through OpenAI’s servers; don’t use for tasks involving banking data, passwords, or sensitive personal information

Common problems and how to solve them

The tools button doesn’t appear in the app

First, verify the app is on the latest version. If yes, check if your account has the compatible plan by accessing account settings on OpenAI’s website. The feature has gradual release and may not have reached your profile yet even with active subscription; in such cases, the wait is usually a few days.

The agent freezes in the middle of execution

Sometimes the agent gets stuck trying to access a website that blocked access. Cancel execution, rewrite the instruction avoiding mentioning that specific website, and ask to search for the information from alternative sources. The agent usually adapts the route when directed.

The result is incomplete or wrong

Add more context in the same conversation. Say specifically what’s missing or what’s incorrect. The agent uses session history, so you don’t need to repeat everything, just correct the specific point.

The app closes during execution

Long tasks consume memory and may be interrupted on phones with less RAM. Close other apps before starting a heavy task. When the app reopens, the conversation history is usually saved, but execution doesn’t continue automatically; you need to ask the agent to resume.

Frequently asked questions about ChatGPT Agent on mobile

Does ChatGPT Agent work without internet?

No. The agent depends entirely on connection because actions are processed on OpenAI’s servers and searches happen on the web in real time. Without internet, you only have access to the basic language model, without agent capabilities.

Do you need to pay extra to use the agent or is it included in Plus?

According to OpenAI’s public information, agent mode is available to Plus, Pro, and Team subscribers at no additional cost, with usage limits varying by plan. The Pro plan (US$200/month) has more generous limits. Confirm current details on OpenAI’s plans page, as the company adjusts policies frequently.

Can the agent make purchases or financial transactions for me?

OpenAI has intentionally limited actions involving real financial transactions. The agent can research options, compare prices, and even access checkout pages, but payment processing is outside the scope for security reasons. Never provide card data in an agent session.

Is there a difference between using it on mobile and on computer?

The core functionality is the same, but the mobile app interface has some layout differences from the desktop browser. Tasks involving file creation or data export may be easier to manage on computer. For queries, searches, and summaries, mobile works just as well.

For whom ChatGPT Agent on mobile makes sense

If you use your phone for work and spend considerable time researching information, comparing options, or producing content, the agent delivers real value. The learning curve is in writing good instructions, but once you get the hang of it, the time savings are notable.

For those who use ChatGPT only for conversation or quick questions, the free plan with the standard model already solves it. The agent makes sense when the task has steps, needs external information, and would take several minutes of your time to do manually.

The feature still has real limitations, especially with Brazilian websites that block automated access, and requires attention when confirming actions. But as a starting point for automating research and information organization right from your phone, it’s one of the most practical features OpenAI has launched so far.