The EU AI Act does not define “AI agent” as a separate legal category. The European Commission says existing rules for AI systems and general-purpose AI (GPAI) models can cover agents in appropriate cases, depending on what a system does and how it is built. The Commission describes its agent-related regulatory considerations as preliminary. Its FAQ explains how the Act’s existing rules may apply.
Treat “agent” as a description of a system, not as its legal classification.
What does the Commission mean by an AI agent?
The Commission describes an agent as a system that can receive and process input from its environment, then act on that processing. Agents typically include a GPAI model and an interface that forms part of an AI system, but the Commission cautions that the term is used inconsistently and does not describe one universal architecture.
That distinction matters because a product’s label alone does not determine which AI Act rules apply. The relevant questions differ by regulatory route: some concern the system and what it is intended to do; another concerns the underlying model.
Which AI Act rules may apply?
The Commission describes several conditional routes. They do not amount to one automatic package of obligations for every system marketed as an agent.
| Regulatory route | What it concerns | When it may apply |
| Prohibited practices under Article 5 | Practices involving harmful manipulation or exploitation of vulnerabilities | The Commission says design safeguards may be needed where prohibited practices are reasonably likely to cause significant harm. |
| Transparency under Article 50 | An agent’s interaction with people or generation of content | The Commission says the rules apply from August 2, 2026, when an agent is intended to interact with natural persons or generate content. |
| High-risk requirements under Chapter III | The agent as an AI system | Additional requirements apply if the system is classified as high-risk. |
| GPAI systemic-risk obligations | The model underlying an agent | Autonomy and tool use may be decisive when assessing whether a GPAI model has systemic risk. Providers of models with that designation have risk-management obligations that consider autonomous capabilities and agentic use. |
The Article 50 condition is tied to the agent’s intended use: interacting with people or generating content. It is not a blanket rule triggered simply by the word “agent.”
How are system-level and model-level risks different?
High-risk classification concerns the AI system; systemic-risk designation concerns the underlying GPAI model. They are separate regulatory questions. An agent’s autonomy or use of tools may matter to the model-level systemic-risk assessment, even though that is distinct from classifying the agent system as high-risk.
In practical terms, examining the system’s intended purpose and the role of its underlying model helps clarify which route is relevant. The Commission’s guidance does not assign a classification to every agent as a category.
The Commission calls its considerations on agents preliminary. Its AI Office monitors developments and may consider further strategies if needed.