Summary Under the proposed Cloud and AI Development Act (CADA), Article 2(5) defines an "AI agent" as "an AI system or a coordinated set of AI systems, that can perceive and act upon their environment, with a degree of autonomy, using tools as needed to achieve specific goals and adapt to changing inputs and contexts." This is an original CADA definition with no equivalent in the AI Act. It singles out autonomous, tool-using, goal-directed systems — and CADA backs it with a dedicated supply-side objective: supporting advanced platforms for the large-scale deployment of AI agents (operational objective 6, in Articles 3 and 4). As CADA is a proposal, the wording could change before adoption.

Detail

CADA introduces a specific defined term for "AI agent" to capture a class of systems that go beyond static models — systems that act, not just output. The definition then connects to a concrete policy objective in the proposal's leadership initiatives.

The definition: Article 2(5)

As proposed, Article 2(5) reads:

"'AI agent' means an AI system or a coordinated set of AI systems, that can perceive and act upon their environment, with a degree of autonomy, using tools as needed to achieve specific goals and adapt to changing inputs and contexts".

Three features distinguish an AI agent from an AI system more generally:

  1. Perception and action. It does not merely process input; it perceives an environment and acts within it.
  2. Autonomy. It operates with some independence from continuous human direction.
  3. Tool use and adaptation. It uses tools as needed to pursue specific goals and adapts to changing inputs and contexts.

Note that it can be a single AI system or a coordinated set of AI systems — multi-agent arrangements are within the definition.

The policy hook: operational objective 6

CADA does not regulate or restrict AI agents through this definition; it promotes them. Article 3(2)(f), as proposed, sets out operational objective 6: "supporting the development of advanced platforms for the large-scale deployment of AI agents." Article 4(6) then details what that entails. Under operational objective 6, the Cloud and AI Leadership Initiatives shall:

  • "support the development of advanced resilient and secure platforms for the development, deployment and orchestration of advanced AI agents at scale"; and
  • "facilitate the development of targeted testing and experimentation methodologies of advanced AI agents and their orchestration throughout their lifecycle."

The recitals reinforce the point: the proposal recognises that as AI agents become more capable, testing and orchestration methodologies are needed to minimise unintended autonomous behaviour. So the policy emphasis is on resilient, secure, well-orchestrated agent platforms — treating agents as infrastructure to be supported and made trustworthy, not as a category to be prohibited.

Why a separate term

Defining AI agents apart from AI systems generally acknowledges that autonomy and tool use raise distinct considerations — operationally and from a security standpoint — compared with passive models. The proposal's focus on secure, resilient platforms signals that the EU wants the infrastructure underpinning autonomous systems to be dependable and aligned with broader EU digital standards.

What this means for you

For CTOs, architects and SMEs, the AI agent definition has practical implications.

1. Product classification. If your system perceives an environment, acts with autonomy and uses tools to pursue goals, it likely meets the Article 2(5) definition. That spans software agents (for example, agents that call APIs and run multi-step workflows) and multi-agent systems, as well as agents embedded in physical systems.

2. Infrastructure expectations. Operational objective 6 (Article 4(6)) centres on platforms for developing, deploying and orchestrating agents at scale. Having a capable model is not enough; expect emphasis on lifecycle orchestration, testing and experimentation frameworks.

3. Security and resilience. The proposal's repeated stress on "resilient and secure" platforms means security will be central to the support and platforms it envisages. Strong identity and access controls, audit trails for autonomous actions, and safeguards against unintended behaviour are sensible priorities.

4. Opportunity. Because agent platforms are a named strategic objective, there is likely to be associated support. SMEs and start-ups with secure, scalable agent platforms may find openings to collaborate within the leadership initiatives.

Common misconceptions

"Every AI system is an AI agent." No. An AI agent must perceive and act on its environment with autonomy and use tools toward goals. A model that only generates text from a prompt, without autonomous action or tool use, is an AI system but not necessarily an agent.

"AI agents means physical robots only." The definition is not limited to physical systems. Software agents that operate in digital environments — calling tools and APIs, automating complex workflows — are covered where they meet the autonomy, perception and action criteria.

"CADA bans or restricts AI agents." On the contrary. The proposal aims to strengthen the EU's ecosystem for AI agents through operational objective 6, supporting secure, resilient platforms rather than imposing prohibitions.

"The AI Act already defines AI agent the same way." It does not. The AI Act defines "AI system" but contains no separate, equivalent defined term for "AI agent". CADA's definition is original to this proposal and reflects a deliberate focus on autonomous, goal-directed systems.

Official sources

Related

This is general information about a draft EU regulation, not legal advice.