The AI Learning Hub
Published 1 November 2026

What an AI Agent Actually Is (and Five Questions to Ask Before You Trust One)

“Agent” has become the most stretched word in AI marketing. It gets used for chatbots, for automations that were around long before AI, and for systems that can genuinely work on their own for hours. If you are being asked to buy, build or approve one, it helps to know what the word should mean and what to ask before you trust it.

The short definition

A chat assistant answers. An agent acts.

More precisely, an AI agent is a language model running in a loop with access to tools. It receives a goal, decides on a next step, uses a tool to take that step, looks at the result, and decides what to do next. It repeats this until it judges the goal is met or it gets stuck.

The tools are what make it an agent. They might let it search the web, read and write files, run code, send email, update a spreadsheet, or click through a website. Without tools, a model can only describe what should happen. With them, it can make things happen.

That loop is also why agents feel so different to use. You don’t approve each sentence; you approve a goal, and the agent chooses the steps. That is the whole point, and it is also where the risk comes from.

Where agents do well

Agents are strongest on tasks that are long and tedious but easy to check. Examples include gathering information from many sources into one summary, fixing a bug where a test shows whether the fix worked, filling in forms from documents, or tidying data. In these tasks the agent saves hours, and a human can verify the result in minutes.

Where agents fail

Their failures are different from a chatbot’s, and usually more expensive:

  • Small errors compound. A model that is right 95% of the time per step is wrong far more often over a 30-step task. Early mistakes get built on rather than caught.
  • They stop too early, or don’t stop at all. Agents can declare success on an unfinished task, or loop on a problem they can’t solve.
  • They follow instructions they shouldn’t. An agent that reads a web page or an email can be tricked by instructions hidden in that content. This is called prompt injection, and it is the most important security problem in agents today.
  • Mistakes are actions, not words. A wrong chatbot answer is a bad paragraph. A wrong agent action is a sent email, a deleted file or a changed record.

The combination to watch for

Security researchers point to one combination that is especially dangerous. It is sometimes called the “lethal trifecta”: an agent that can

  1. read private data,
  2. take in untrusted content (web pages, emails, documents from outsiders), and
  3. send information out (email, web requests, posting messages).

With all three, a single malicious document can instruct the agent to collect private data and send it to an attacker, and the agent may do it while appearing to work normally. Removing any one of the three breaks the attack. That makes it a useful first check for any agent design.

Five questions to ask before you trust one

Whether a vendor is selling you an agent or your own team is building one, ask:

  1. What tools can it use, and what is each one allowed to do? “It has access to email” should become “it can read the support inbox and draft replies, but not send them.”
  2. Does it have the lethal trifecta? If it reads private data and untrusted content and can send things out, ask how that is contained.
  3. Which actions need a human to approve them? Anything hard to undo, such as payments, deletions, external messages or changes to records, should pause for a person.
  4. What is the worst thing it could do in one run? Then ask how that limit is enforced: by permissions the agent can’t change, or just by asking it nicely in its instructions?
  5. How do you know it worked? Is there a log of every action it took, and a test showing how often it completes the task correctly?

A good answer to all five doesn’t make an agent safe, but a vague answer to any of them is a reason to wait.

Where to go deeper

Agent Engineering covers how agents are built, from the tool loop to evaluation, for engineers shipping them. Securing AI Systems goes deep on prompt injection and the lethal trifecta, and our threat-model canvas gives you a one-page way to apply these questions to a real feature. For the basics of how the underlying models work, start with AI Essentials. Free, like everything on the AI Learning Hub.

Related free courses