Is ChatGPT an AI agent? Every mode scored against five tests

Is ChatGPT an AI agent? Plain chat is not; agent mode mostly is. Each ChatGPT mode scored on five tests, one task run two ways, and the EU rules that apply.

Published
Reading
7 min
By
Authoricy

ChatGPT is an AI agent only when you switch on agent mode, which OpenAI launched on 17 July 2025. Plain ChatGPT, the chat window you type into, is an AI assistant: it answers and then waits for you. Agent mode takes a goal, uses a browser and other tools, and carries out the steps itself.

So the honest answer is “it depends which ChatGPT”. Search, deep research and custom GPTs each land somewhere in between. If you already know you need an agent that works inside your own systems, that is what our AI agent development team builds.

Five tests for an AI agent

There is no legal definition of an AI agent, but the companies building them agree on the core. OpenAI calls agents systems that independently accomplish tasks on behalf of users. Anthropic says agents are systems where the model directs its own process and tool use, working in a loop and reading the real result of each step before the next.

Turn those into tests you can apply to any product:

  1. Goal. It works toward an outcome you set, not a single reply.
  2. Tools. It can use something beyond its own text: a browser, code, files, an API.
  3. Loop. It plans, acts, reads the result and decides the next step without a new prompt each time.
  4. Action. It changes something outside the chat: sends, books, files, updates.
  5. Control. It checks its own work and stops or hands over when something is wrong.

A chatbot passes none or one. An assistant passes two or three. An agent passes all five. Our guide to AI agents for business covers how that difference plays out in daily operations.

Every ChatGPT mode, scored

ChatGPT is not one product. The same app holds several modes, and each one lands in a different place.

Mode Goal Tools Loop Action Control Verdict
Chat No No No No No Assistant
Search No Web search No No No Assistant with sources
Deep research Yes Browser Yes No Partial Research agent, read-only
Custom GPT with actions Partial Your APIs Partial Yes Partial Assistant that can act when asked
Agent mode Yes Browser, terminal, connectors Yes Yes Yes, with confirmations Agent
API with the Agents SDK Yes Whatever you build Yes Yes Whatever you build Agent you design

Two rows need comment. Deep research runs a real loop. It searches, reads, follows leads and writes a report, but it only reads, so it fails the action test. A custom GPT with actions can call your systems, but each call usually follows one user message rather than a plan it made itself.

The last row is the developer route. OpenAI ships an open-source Agents SDK with handoffs, guardrails and tracing. With it you can build an agent on the same models that never opens a chat window at all.

What ChatGPT agent mode does, and where it stops

OpenAI’s launch post for ChatGPT agent describes a merger of two earlier tools. Operator browsed and clicked on websites. Deep research read and summarised. Agent mode puts both inside ChatGPT, running on a virtual computer of its own.

What it can do, according to OpenAI:

  • browse and fill in websites, and log in when you take over the browser
  • use connectors such as Gmail and GitHub to read your data
  • run code and produce files such as spreadsheets and slides
  • repeat a finished task on a schedule, for example a weekly report

Where it stops is just as important for a business:

  • Before a purchase or any other “action of consequence”, ChatGPT requests permission.
  • Some tasks, such as sending email, need you watching while it runs.
  • High-risk tasks such as bank transfers are refused by training.
  • OpenAI names prompt injection as a risk: a web page can hide instructions meant to hijack the agent.

Those limits mean agent mode is built for a person beside it, not for unattended work at 06:00.

One task, done two ways

Take a job most sales teams know. Every Monday, someone finds the quotes in the CRM that have waited a week without an answer. Each customer gets a short follow-up, and the CRM gets a note that it went out.

With ChatGPT agent mode

  1. A seller opens ChatGPT, describes the task and starts agent mode.
  2. The agent opens the CRM in its browser. The seller takes over to log in.
  3. Next it reads the list of open quotes and drafts a follow-up for each.
  4. Before it sends, it asks. Sending email runs under the seller’s watch.
  5. Logging each follow-up means more CRM clicks, each depending on an unchanged page.
  6. The seller can schedule a repeat, but someone still has to be there for the login and the send.

For twenty quotes a month this may be enough. The cost is a person’s attention on every run, with no record outside the chat.

With a purpose-built agent

  1. The agent starts on Monday morning by itself.
  2. Open quotes come from the CRM’s API rather than its screen, so a redesign does not break the run.
  3. It applies your rules. It skips customers who replied by phone, never sends a third reminder, and holds anything above a set deal size for the seller.
  4. Each follow-up goes from the seller’s own address and writes the follow-up into the customer’s record.
  5. Every step is logged: what it read, what it decided, what it sent.
  6. Exceptions land on a named person’s desk with a one-line reason.

Here the person reviews exceptions instead of doing the run. That pays off only when the volume is high and the rules are clear. Our list of AI agent examples shows which other jobs fit the same shape.

When ChatGPT is enough, and when it is not

The more “yes” answers below, the stronger the case for your own agent.

  1. Does the task repeat every day or every week, at a volume that eats hours?
  2. Should it run when nobody is at a screen?
  3. Does it have to write into your ERP, CRM or ticket system, not just read from them?
  4. Do you need a record of every decision, for a customer, an auditor or your own team?
  5. Are there business rules a prompt cannot hold reliably, such as limits, approvals and exceptions?
  6. Does it handle customer data that must stay inside systems and contracts you control?

If most answers are “no”, use ChatGPT, or start with a plain rule-based flow. Our guide to AI workflow automation covers when that simpler route is the right one. If most are “yes”, the task needs an agent built around your systems and rules. AI agent development shows how one gets built, step by step.

Phone calls differ: a caller expects an answer in seconds, so an AI receptionist for a small business is built for live voice, which a chat tool is not.

The European layer

Three things change for a company in the EU.

Availability. At launch, OpenAI said it was “still working on enabling access for the European Economic Area and Switzerland”. Access depends on region and plan, so check what your own account offers before you plan around it.

Transparency. Article 50 of the EU AI Act says AI systems that interact directly with people must be built so those people are informed they are interacting with an AI system. The exception is when this is already obvious. An agent writing to customers in a seller’s name needs this thought through.

GDPR. When an agent logs into your mailbox or CRM, it processes personal data. Under the GDPR that needs a lawful basis (Article 6), a data processing agreement with the provider (Article 28) and a clear answer to where the data goes. Article 22 limits decisions based solely on automated processing that have legal or similarly significant effects on a person. Credit, pricing or cancelling a customer are examples, so keep a person in the loop for those.

None of this rules agents out, but the questions about logging, hand-over and data location belong in the design from day one. That is why every project with our AI agent development team in Stockholm starts with a fixed-price discovery. It ends in a written blueprint and a fixed price for the build, and you own the code.

Frequently asked questions

Is ChatGPT an AI agent or an LLM?

Both words apply to different layers. The model underneath is a large language model. ChatGPT in plain chat mode is an assistant built on that model. Agent mode wraps the same models in tools and a loop, which makes it an agent.

What is the difference between an AI assistant and an AI agent?

An assistant answers when you ask and leaves the next step to you. An agent takes a goal, decides the steps, uses tools to carry them out and changes things in other systems. The test is simple: when the chat ends, has anything happened outside it?

Can ChatGPT take actions on its own?

In agent mode, yes, within limits. It browses, fills in forms and works with connected apps. OpenAI has it ask before consequential actions, keep you watching while it sends email, and refuse high-risk tasks such as bank transfers. In plain chat it takes no actions at all.

Can I build my own AI agent with ChatGPT?

Not inside the chat app in the full sense. A custom GPT with actions can call your systems. An agent that runs on its own schedule, follows hard rules and logs every step is built on OpenAI’s API and Agents SDK, or on another model provider. That is a software project, with design, testing and monitoring like any other.

Talk to the people who build it

Name the task your team dreads.

Book a call and we tell you honestly whether an agent should do it, and what it would take.