AI agents, formerly also referred to as intelligent agents, are computer systems or programmes that can perform tasks autonomously on behalf of the user or another system by organising its workflow and using available tools. The functionalities of these agents go beyond traditional automation tools, as they not only follow a set of instructions, but can also think, adapt and act independently.

AI agents still need human supervision as we have not yet discovered artificial general intelligence. Current AI models are still capable of only performing tasks well and are not intelligent per se.
The agents however will help humans focus on tasks that require higher cognitive skills by saving them time.

RPA was done using a fixed set of rules and hence was not flexible enough to handle different scenarios intelligently.
AI agents on the other handle have a bit of intelligence and are therefore able to adapt to changing situations on the fly.
AI agents can also provide a natural language interface to the users so that they do not have to learn special commands to operate.

AI agents rely on large language models (LLMs) for their intelligence. These models are intelligent, however are unpredictable and sometimes suffer from a phenomenon called hallucination.
Hence human supervision is always recommended, especially when the AI agent is deployed in a customer facing app.
We can also add guardrails, checks and balances to lower the risk of agents behaving unpredictably. One such technique is to assign one agent to monitor another agent.