Skip to main content

EU AI Act rules can cover agents under existing definitions

21 NOVEMBER 2026·3 MIN READ·4 SOURCES

A Unite.AI article published November 19, 2026, says Muller and Christiansen argue that agentic AI is already covered by the EU AI Act, rather than occupying a regulatory gap. The European Commission says the Act has no separate “AI agent” category, but its existing definitions can cover agents.

Key takeaways · 4

  • 01

    The Commission does not treat “AI agent” as a separate category; assess whether the system and its underlying model meet the Act’s existing definitions.

  • 02

    An agent is described as receiving and processing environmental input, then taking actions based on that processing.

  • 03

    Transparency rules apply from August 2, 2026, when an agent is intended to interact with people or generate content.

  • 04

    High-risk agents face additional requirements from December 2, 2027, or August 2, 2028, depending on the applicable transition period.

Existing definitions, not a new category

The European Commission says “AI agent” is not a separate category under the AI Act, but the Act’s definitions of AI systems and GPAI models can cover agents.[2] It describes an agent as able to receive and process environmental input, then take actions based on that processing.[2] The Commission adds that an agent typically contains at least one GPAI model and usually has an interface that forms part of an AI system.[2] It also notes that “AI agent” is used inconsistently in public debate.[2] That makes the specific system’s design and functions more useful for assessment than the label alone.[1]

Obligations depend on use and risk

The Commission says rules for AI systems and GPAI models also apply to agents that meet the relevant definitions.[2] The Act is described by the Commission as a risk-based framework in which rules for developers and deployers depend on specific uses of AI.[4] The Commission identifies four levels of risk for AI systems.[4] The Unite.AI authors argue that providers of systems that could affect people’s rights should establish comprehensive risk-management systems, including assessing possible user misuse.[1] They say the Act focuses on health, safety, and fundamental rights.[1] These are the authors’ arguments about the Act’s implications, distinct from the Commission’s FAQ guidance.

Dates differ by obligation

From August 2, 2026, transparency rules apply when an agent is intended to interact with natural persons or generate content, according to the Commission.[2] If an agent is classified as a high-risk AI system, additional requirements apply from December 2, 2027, or August 2, 2028, depending on the relevant transition period.[2] The dates attach to different obligations, so teams assessing deployment should avoid treating them as a single deadline. The evidence does not specify which transition period applies to a particular agent; determining that requires assessing the system’s circumstances against the applicable rules.[2]

Agent and model risks are distinct

A high-risk classification concerns the agent as an AI system, while systemic-risk designation concerns the underlying GPAI model, according to a NeoTeo recap.[3] The Commission says the underlying model’s autonomy and tool use can be decisive in whether it is designated as a model with systemic risk.[2] Providers of GPAI models designated as systemic risk have risk-management obligations that include considering autonomous capabilities and agentic use.[2] The Commission calls its regulatory considerations about agents preliminary; it says the AI Office is monitoring developments and may develop strategies if needed.[2] Its AI safety technical-assistance tender also includes work evaluating agent safety and security.[2]

Teams evaluating AI agents should look beyond product labels and document what systems do, how they are used, and which risks their functions may raise. Separating system-level high-risk classification from model-level systemic-risk designation can help direct reviews to the relevant obligations.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 21 November 2026

    EU AI Act rules can cover agents under existing definitions

Sources

AI fluency, one session a day, built for your work.