Agentic AI
Agentic AI, or an AI agent, is an artificial intelligence system that can not only analyse or generate content but also act: plan tasks, call tools, send messages and chain steps together without intermediate approval.
What changes compared with a generative model
Generative AI produces content that a person reads before doing anything with it. An agent executes. The consequence concerns the nature of the risk: a wrong answer becomes a wrong action, sometimes an irreversible one.
The design therefore has to change:
- human oversight must be placed before the action, not after it;
- the permissions granted to the agent define how much damage is possible: an agent with write access can destroy what it was meant to organise;
- traceability of every action becomes essential to reconstruct a sequence of events afterwards.
The AI Act classifies these uses according to their level of risk; those involving decisions about people generally fall into the high-risk category.
Environmental issue
An agent multiplies calls to the model: every reasoning step, every tool invoked and every failed attempt consumes an inference. The energy cost of an agentic task is far greater than that of a single request.
