AI · explainer
What an AI agent actually is — and what it is not
The word ‘agent’ is being stretched until it means everything. Here is a practical definition newsrooms and buyers can use.
Key takeaways
- An agent is software that can plan and use tools toward a goal, not just answer a prompt.
- Most products labeled ‘agentic’ are still assisted workflows with a human in the loop.
- Risk rises when an agent can take irreversible actions without a review step.
The useful definition
An AI agent is a system that can take a goal, break it into steps, call tools, observe the result, and continue until the goal is met or it must stop.
A chatbot answers. A copilot helps a human finish a task. An agent is allowed to act.
If the system cannot use tools, cannot persist state, and cannot fail safely, it is not an agent. It is a pipeline with marketing.
Why the confusion happened
Language models got good at sounding decisive. In production, most teams still insert a human approval step the moment money, email, or production infrastructure is involved.
What to ask a vendor
- Which tools can it call, and can you revoke them?
- What is logged — prompts, tool output, and who approved the action?
- What happens when evidence is missing: stop, or invent?
The Intel Brief rule
Treat agent claims like any other factual claim. Demand a primary source, a demo you can reproduce, and a named owner when something goes wrong.