Every James Bond movie starts the same way.
Bond gets a mission. Q hands him a few gadgets. Bond heads out and figures out the rest on his own. Meanwhile, M waits back at headquarters hoping nothing blows up.
(Spoiler alert, something always blows up).
It turns out that James Bond is a pretty good analogy of how an AI agent works.
You give an AI agent a goal. It gather the tools, like your browser, your files and your apps. Then it works out the steps by itself and reports back when the job’s done. (Ideally with no explosions.)
The Confusing Rise of AI Agents
AI agents are the biggest change in how we use AI since ChatGPT arrived in 2022. They’re also one of the most confusing tools in tech right now.
In the US, the question “what is agentic AI” gets searched around 33,000 times a month on Google alone. But there is a lot of worry, and misinformation about what an AI agent is, and how it works.
In this article, I’ll explain what AI agents are in plain English, how they work, and what they can and can’t do yet.
We’ll also cover the risks of agentic AI worth knowing about before you hand one the keys.
What Is an AI Agent?
Short answer: An AI agent is AI that can plan the steps and take actions for you. You set the task. It plans and does the work. You check.
IBM’s explainer on what AI agents are describes them as software that carries out tasks on its own by mapping out a workflow and using the tools it has access to.
The word itself is a clue. “Agent” comes from the Latin agere, which means to drive, lead or act. An agent is something that does things.
That’s the big difference from a regular chatbot. A chatbot gives you an answer, and you still do the work. An AI agent takes the next step for you.
What Is Agentic AI?
Short answer: Agentic AI is the umbrella term for AI that can act on its own. “Agentic” describes the ability. An AI agent is the thing that has it.
In practice, agentic AI usually has four traits:
- It can act independently. It doesn’t need a new prompt for every step.
- It chooses its own next steps. You give it the goal, and it decides the route.
- It has access to files and logins. That’s what lets it actually do things.
- It checks in before and after. The good ones do, anyway.
Anthropic draws a helpful line between two kinds of AI systems in its guide to building effective agents. In a workflow, people set the steps in advance. With an agent, the AI decides its own steps and which tools to use.
One heads-up: “agentic” has become a very popular sticker. Gartner calls it “agent washing,” where chatbots and old automation tools get relabelled as agents. Gartner estimates only about 130 of the thousands of companies selling agentic AI are offering the real thing.
Is ChatGPT an AI Agent?
Short answer: A regular ChatGPT conversation isn’t an agent. But ChatGPT, Claude and Gemini all now have agent features built in.
When you chat with ChatGPT or Claude, you ask and it answers. You still copy, paste, click and send.
The agent features are different. ChatGPT agent can browse websites and complete tasks for you. As of last week, ChatGPT can also work inside your open Chrome tabs, filling in forms and downloading files. On the Claude side, Claude in Chrome does the same kind of browser work, Claude Code writes and fixes code, and Claude Cowork organizes files and builds documents on your computer.
While I was writing this article, my AI co-host Shorty (who runs on Claude) opened SEMrush in my Chrome browser, searched for the prompts people ask AI about agents, and pulled the data that shaped the headings you’re reading. I watched it click around like a very keen intern.
That’s agentic AI in one sentence. I set the task, and it did the clicking.
How Do AI Agents Work?
Short answer: AI agents work in a loop. They make a plan, use tools to take action, check the result, then repeat until the job is done.
- Make a plan. The agent breaks your goal into small tasks.
- Use tools. It searches, clicks, types, sends and saves, using the apps and logins you’ve given it.
- Check, then repeat. It looks at what happened and tries again until the task is finished.
IBM describes the same idea as three parts working together: planning, tools and memory. Memory is what lets an agent remember what it already tried, so it doesn’t repeat the same mistake.
The loop is where the magic happens. It’s also where things go wrong. A small error in step two can snowball by step ten, and every extra lap of the loop uses more AI tokens. That’s one reason agent tasks can eat through your usage limits so fast.
What Are Some Examples of AI Agents?
Short answer: Today’s best-known AI agents include ChatGPT agent, Claude in Chrome, Claude Code, OpenAI’s Codex and Claude Cowork. But the idea of an AI agent is much older than ChatGPT.
Here’s a quick history of AI agents:
- 1966-1972: Shakey the robot. Built at the Stanford Research Institute, Shakey could take a goal, make a plan, move around a room and check its progress. Life magazine called it “the first electronic person.”
- 1997: Clippy. Microsoft’s Office assistant could suggest things, but it couldn’t do anything for you.
- 2011 and 2014: Siri and Alexa. Voice assistants that could set timers and play music, as long as the task followed a script. (We asked whether the new Siri still sucks in Episode 1.)
- 2023: AutoGPT. A free tool that let GPT-4 write its own to-do list and work through it. It was clumsy and often got stuck, but millions of people saw what an AI agent could be.
- 2024-2025: Agents get ‘hands’. Anthropic released “computer use” in October 2024, letting Claude move a mouse and type. OpenAI followed with Operator in January 2025 and ChatGPT agent in July 2025.
So the AI agent is actually about 60 years old. What’s new is the brain inside it. Shakey needed a room full of computers to push a box across the floor. Today’s AI agents can work across the whole internet without a human being involved.
What Are the Best Use Cases for AI Agents?
Short answer: AI agents work best on tasks with lots of steps, clear goals and a human checking the result. Research, writing, coding, file work and planning are the strongest use cases right now.
Here’s what AI agents can do well today:
- Research and write content
- Search through websites for you
- Write and fix code (hello, vibe coding)
- Sort files and build documents
- Keep you on track with plans
- Brainstorm new ideas
- Look for patterns and trends
- Track results and refine strategy
For the show, Shorty does a lot of this list every week. It researches the topic, checks the sources and builds a first draft of the slides. (I still rework them heavily before we go live.)
AI agents are also improving quickly. Research group METR measured how long a task AI agents can finish, and found that length has doubled about every seven months for six years. That was measured on fairly tidy software tasks, so real office work is messier. But the direction is clear.
What Can’t AI Agents Do Yet?
Short answer: AI agents still struggle with messy real-world work. They get stuck on surprises, get talked into bad decisions, and don’t always know when they’re wrong.
Two experiments show the gap nicely.
Researchers at Carnegie Mellon built a fake software company called TheAgentCompany and staffed it entirely with AI agents. The best performer, Claude, finished about 24% of its tasks. One agent hit a pop-up window it couldn’t close and gave up. A human would have just clicked the little X.
Anthropic tried something similar with Project Vend, where Claude ran a small office shop for a month. It found rare snacks and refused sketchy requests. It also got talked into discount codes, sold tungsten cubes at a loss, and told customers to pay into a Venmo account that didn’t exist. At one point it announced it would deliver orders in person, wearing a blue blazer and a red tie.
The last item on the list matters most. We covered why AI makes things up in Episode 3. AI agents can hallucinate too. The difference is that an agent might act on the hallucination.
What Are the Risks of Agentic AI?
Short answer: The biggest risks of agentic AI are rogue AI agents that break the rules you set, and shadow AI, where people use AI tools at work without approval. Agents can also reply as you to other people, and spend money or make choices for you.
Rogue AI agents
A rogue AI agent is one that goes outside its instructions or permissions.
In July 2025, SaaStr founder Jason Lemkin was vibe coding with Replit’s AI agent. During a code freeze, with clear instructions to make no changes, the agent deleted his live database. In the chat afterwards, it admitted it had “panicked” and “violated explicit instructions.” Replit’s CEO called it unacceptable and added new safeguards.
Agents don’t only work for the good guys, either. This week, Reuters reported that one hacker used AI agents to break into more than 100 company websites, for about $5,000 in AI costs.
Shadow AI
Shadow AI is when people use AI tools at work that the company hasn’t approved. Think client data pasted into a personal chatbot, or an agent connected to a work inbox without anyone in IT knowing.
IBM’s 2025 Cost of a Data Breach report found one in five organizations had a breach linked to shadow AI. Those breaches cost about $670,000 more on average. Only 37% of organizations had policies to manage it.
Here’s the simplest way I can put it. A chatbot can only say something wrong. An AI agent with your logins can do something wrong.
How Do You Create an AI Agent?
Short answer: You don’t need to code to start with AI agents. Most people begin with a ready-made agent like ChatGPT agent, Claude in Chrome or Claude Cowork, then build their own with no-code tools once they know what they want it to do.
Three habits make the first few weeks much smoother:
- Start small and simple. Pick a task you’d happily hand off, like sorting a folder or researching a topic.
- Watch it the first time. You’ll learn a lot from seeing how it plans and where it trips.
- Draft, check, approve. Let the agent draft, but keep the final click on anything that publishes, sends or buys.
If you want to build your own, the line between workflows and agents is a good place to start. Many tasks work better as a fixed AI workflow automation with set steps, and only need a full AI agent when the steps change every time.
Every Good AI Agent Needs a Handler
Even James Bond has M.
The best AI agents in 2026 are fast, capable and getting better every few months. They still need someone to set the mission, hand over the right gadgets, and check the results before anything gets sent.
That’s the human job now. Less doing every step yourself, more deciding what’s worth doing and checking that it got done well.
And if your agent ever offers to deliver something in person wearing a blue blazer, it’s probably time for a debrief.
– Daniel Midson-Short
Leave a Reply