Summary Under the proposed Cloud and AI Development Act (CADA), an AI agent is distinguished from an ordinary AI system by its ability to perceive and act on its environment, with a degree of autonomy, using tools as needed to pursue goals and adapt to changing context. An ordinary "AI system" (Article 2(3), via the AI Act) is defined broadly by inferring outputs from inputs; an "AI agent" (Article 2(5)) layers on perception, action, autonomous goal-seeking, tool use and adaptability — possibly across a coordinated set of systems. CADA singles out agents for dedicated support because their autonomous-execution paradigm needs orchestration, testing and security foundations that static models do not.

Detail

The distinction sits in Article 2 of the proposal, which separates passive inference from active, autonomous action.

The definition of an AI system

Article 2(3) defines an "AI system" by reference to Article 3, point (1), of Regulation (EU) 2024/1689 (the AI Act). That definition describes a machine-based system that operates with varying levels of autonomy and infers, from the input it receives, how to generate outputs — predictions, content, recommendations or decisions — that can influence physical or virtual environments.

This is broad and foundational. A classifier, a predictive-maintenance model or a generative text model all fit. The defining feature is inference: input in, output out. It does not, by itself, require the system to act on that output in the external world or to pursue a goal over time.

The definition of an AI agent

Article 2(5) provides a narrower definition for an "AI agent":

"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."

Four elements lift an AI system to agent status:

  1. Perception and action. An agent perceives its environment and acts upon it, rather than merely emitting an output for a human to use — a closed loop of engagement with the world.
  2. Goal-directed autonomy. While the general definition mentions "varying levels of autonomy," the agent definition ties autonomy to achieving specific goals, with the agent making sequential decisions toward an objective.
  3. Tool use. The agent uses "tools as needed" — calling APIs, running code, controlling actuators, querying databases — which distinguishes it from a self-contained model and shapes its risk profile.
  4. Coordinated sets and adaptability. An agent may be one system or a "coordinated set of AI systems," reflecting that agentic behaviour often emerges from orchestrating multiple models, and it must "adapt to changing inputs and contexts."

Why CADA singles out agents for dedicated support

The distinction drives policy in the proposal. Recital 21 notes that as AI agents have become increasingly capable, industry is moving toward equipping systems with autonomous-execution capabilities — and that this calls for robust engineering around orchestration, testing and accountability rather than model accuracy alone.

CADA reflects this in its operational objectives for the Cloud and AI Leadership Initiatives. Article 4(6) (operational objective 6) provides that the 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 for advanced AI agents and their orchestration throughout their lifecycle.

In addition, Annex I, Grand Challenge 7 ("AI Agents Platform") is dedicated to developing a European AI agent orchestration framework — exploring paradigms that let multiple agents collaborate while maintaining rigorous security. The aim is sovereign, cloud-based open platforms for managing AI agents at scale.

By singling out agents, CADA recognises that infrastructure suited to a static model is insufficient for the dynamic, tool-using, autonomous nature of agents, and aims to foster the orchestration and middleware layers that agentic workflows require.

What this means for you

For CTOs, architects and SMEs, the distinction shapes both planning and positioning.

  • Infrastructure planning. If you deploy AI that only returns recommendations or generated content, you are likely in ordinary AI-system territory. If you build solutions that autonomously execute tasks, call external tools, or coordinate multiple models toward complex goals, you are deploying AI agents — the area CADA earmarks for investment and support.
  • Security and accountability. Autonomous execution introduces distinct risks: an agent can take real-world actions with consequences. CADA's emphasis on testing and experimentation methodologies for agents points to scrutiny of orchestration layers and safety guardrails, so design for observability, human-in-the-loop controls and robust error handling.
  • Strategic alignment. If your solution fits the Article 2(5) agent definition, align your roadmap with the operational objective 6 goals (Article 4(6)) and the European AI agent orchestration framework (Annex I, Grand Challenge 7) to position for support within the EU's cloud ecosystem.

Common misconceptions

  • "All autonomous AI systems are agents." Not necessarily. A system can act autonomously in a narrow sense (for example auto-scaling server load) without meeting Article 2(5) if it does not perceive a broader environment, use tools to pursue goals, and adapt to changing context. The combination matters.
  • "An AI agent is just a chatbot with API access." A chatbot can be an agent if it meets the criteria, but a chatbot that only retrieves and displays information is an AI system. It becomes an agent when it can plan, use tools to execute tasks, and adapt toward a goal beyond the immediate conversational turn.
  • "CADA regulates agents more strictly than other AI systems." As proposed, CADA is a development and capacity-building act, not a risk-based regime like the AI Act. It does not impose stricter penalties on agents; it provides targeted support and infrastructure because agents are seen as strategically important and technically demanding.

Official sources

Related

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