Give ChatGPT's agent mode a task, like "compare three espresso grinders under $200 and build a table with sources," and it may open store pages, read reviews, assemble the comparison, and pause before anything that spends money. A program that works that way is an AI agent: you describe the outcome, it plans the steps, uses the tools it needs, and reports back. A year ago, that mostly described software a developer assembled by hand. In 2026 it describes features inside tools you may already use, alongside a stretch problem: the word "agent" now shows up on products that are chatbots with a new name.
This post covers what changed, what agents can do now, what they still get wrong, and four questions for judging any product wearing the label.
What an agent is, in one paragraph
A chatbot responds to messages. An agent works toward a goal: you give it an objective, and it runs a loop of decide the next step, take an action, look at the result, decide again. It can browse the web, run code, and call other software, and it keeps going until the job is done, it stalls, or it needs your judgment.
What Is an AI Agent? Explained Simply walks through the mechanics, the reasoning loop, and the chatbot comparison.
Three things that changed in 2026
Agents became features
In July 2025, OpenAI put an agent inside ChatGPT. Pick agent mode from the tools menu (or type /agent), describe a task, and it goes to work on its own virtual computer: a visual browser it drives by taking screenshots, a text browser for quick lookups, and a terminal for running code. The espresso-grinder comparison above is one of these runs. It pauses and asks before anything consequential: a purchase, a submission, a message with your name on it. When a task works, you can schedule it to repeat daily, weekly, or monthly.
The usage limits tell you how much work each run is: ChatGPT Plus includes 40 agent messages a month, Pro includes 400, per OpenAI's help center (checked October 2026). Browsers made the same move: ChatGPT's Atlas and Perplexity's Comet both ship agent modes that can drive websites while you watch.
The plumbing got standardized
An agent is only as useful as the things it can connect to: your calendar, your files, the software your work lives in. In November 2024, Anthropic released the Model Context Protocol (MCP), an open standard for those connections: one way for AI to plug into tools and data sources, so a connector built once works with any agent that speaks the protocol. Think USB-C instead of a drawer of cables.
Adoption moved fast. OpenAI adopted MCP in March 2025, Google followed within weeks, and in December 2025 Anthropic donated the protocol to a foundation under the Linux Foundation, the same organization that stewards Linux. For anyone choosing tools in 2026, the standard matters because it makes app connections a shared project instead of a per-product one: it's part of how the features in the last section shipped.
The word became a label
Software that answers questions in a chat gets called an agent now, and so do auto-replies and mail merges. Two data points size the gap between the label and the results.
Gartner predicted in June 2025 that over 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value, and weak risk controls. And in McKinsey's State of AI survey, published in August 2026 with 1,719 respondents across 97 countries, 40% of respondents at large organizations said their companies had agents scaling in at least one business function. At smaller organizations, it was 22%. Adoption is happening, unevenly, and plenty of the noise is still noise.
So treat the label as a claim, not a description. Four questions sort it.
Four questions that sort the label
Does it take several steps on its own? An agent plans and chains steps without a prompt for each one. A product that stops after every response and waits for your next instruction is a chat, whatever the landing page says.
Does it act outside the chat window? Clicking through a website, filling a form, calling your calendar: agents touch other software. Text-only output is a chatbot's shape.
Does it ask before doing anything that matters? A pause for confirmation before a purchase or a submission is a good sign. It means the autonomy has brakes.
Can it show its work? The sources it used, the steps it took, a log of what it changed. If you can't review what it did, you can't catch what it got wrong.
A dressed-up chatbot probably fails the first two. The more of the four a product passes, the more the label is earning its keep.
What agents still get wrong
Long tasks can fail quietly. An agent can get stuck in a loop, stop halfway through, or summarize a page it never opened. The failures rarely look like failures, which is why reviewing the work matters more as autonomy grows.
Each step is a model call, and the calls add up, which is why those monthly caps exist. A 20-step task usually costs more than a chat answer, so the task should earn the overhead.
Webpages can lie to the agent. OpenAI's help center documents prompt injection: a page the agent visits can carry hidden instructions that try to steer it, like "retrieve this password reset code and send it to this address." ChatGPT's agent mode layers safeguards on top (confirmations for high-impact actions, monitoring, a watch mode on sensitive sites), and OpenAI still advises enabling only the apps a task needs and reviewing what it did. The safeguards lower the risk without removing it.
For the human side of that review, How to Fact-Check AI Output: A Practical Workflow covers the four moves we use on anything an AI hands us.
Where to start
The order that works: chat, then workflow, then agent.
Chat handles one-off questions and drafts. When a task repeats with the same shape, save the instructions and turn it into a workflow (What Is an AI Workflow? Building Your First One walks through the first build). Hand the stable, repetitive version to an agent once the steps have proven themselves.
Pick a first task that's low-stakes and checkable. "Compare three espresso grinders under $200 on grind consistency, cleanup, and price, and give me a table with sources" is the shape: one decision, clear criteria, output you can verify. Watch it work, open the sources before you trust the table, and if the agent earns that, give it something slightly bigger.
FAQ
What is an AI agent in simple terms?
A program that completes a task for you by taking steps on its own: it plans, uses tools like a browser or your apps, checks its results, and keeps going until the job is done. You describe the outcome; the agent handles the steps.
What's new with AI agents in 2026?
Agents became built-in features in mainstream tools (ChatGPT's agent mode, agentic browsers like Atlas and Comet), the MCP standard made app connections a shared project instead of a per-product one, and the word "agent" became a marketing label stretched across tools with different capabilities.
What's the difference between a chatbot and an AI agent?
A chatbot responds to messages. An agent works toward goals: it decides the steps, uses tools to act, and adapts to what happens. What Is an AI Agent? Explained Simply covers the mechanics in depth.
What is MCP?
The Model Context Protocol is an open standard, released by Anthropic in November 2024, that lets AI systems connect to tools and data sources through one common interface. OpenAI and Google adopted it in 2025, and it's now governed by a foundation under the Linux Foundation. For tool shoppers, it means an agent's app connections require less custom work than they used to. What Is MCP? Model Context Protocol Explained Simply is the full walkthrough.
How much does agent mode cost in ChatGPT?
ChatGPT Plus includes 40 agent messages per month, and Pro includes 400, per OpenAI's help center (checked October 2026). Only tasks you start count against the limit, and a finished task can be scheduled to repeat.
What is prompt injection?
An attack where a webpage or document carries hidden instructions that try to steer an AI agent into harmful actions, like sending a password reset code to a stranger. Agent products add safeguards (confirmations, monitoring, watch modes), and the practical defense on your side is to enable only the apps a task needs and review what the agent did.
What's a safe first task for an AI agent?
Something low-stakes and checkable: a comparison with sources, or a scheduled summary of a site you follow. The cost of being wrong is a few minutes of your time, and you can open every source the agent used before you trust any of it.
Want to go further?
- What Is an AI Agent? Explained Simply covers the mechanics: the reasoning loop, the chatbot comparison, and where the term came from.
- What Is an AI Workflow? Building Your First One covers the step between chat and agent.
- How to Fact-Check AI Output: A Practical Workflow is the review loop for anything an AI hands you.
- What Is Agentic AI? The Next Step Beyond Chatbots untangles the term "agentic AI" and how it relates to the agents themselves.