Instead of only answering a question, an agent can take a goal such as handling a support ticket, writing and testing code or researching a topic, and work through the steps. It uses tools like search, databases and business software, checks its results and continues until the task is done or it is stopped.
In company results and investor talks, agents are often presented as the next source of revenue for software and cloud firms. Vendors sell them as products, add them to existing subscriptions or charge by usage. The promise is that they automate whole tasks and not only single answers.
The claims need caution. Agents make mistakes that can build up over many steps, they can be manipulated by content they read, and many products described this way offer limited autonomy. Real value depends on reliability and on how much human review is needed.
When you evaluate an agent product, ask what it can actually do, which actions need approval and how success is measured.