Summary Under the proposed Cloud and AI Development Act (CADA), a multi-agent system is not automatically a single monolith β€” but Article 2(5) defines an "AI agent" to include "a coordinated set of AI systems." So an orchestrated multi-agent architecture, where the systems perceive their environment, act with a degree of autonomy and use tools to achieve goals, would itself count as an AI agent. Loosely coupled tools that lack that coordinated, autonomous interaction would fall outside the definition.

Detail

CADA (a proposal, not yet in force) gives the "AI agent" a functional definition that directly addresses modern multi-agent design. For CTOs and architects, the distinction between a single agent and a coordinated set is what would determine scope.

The definition: "a coordinated set"

CADA does not limit an AI agent to a single model or instance. 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"

By including "a coordinated set of AI systems," the proposal would bring multi-agent architectures into the concept. The focus is on the coordinated, autonomous behaviour of the set, not the number of underlying models. If multiple AI systems work together β€” perceiving an environment, exercising autonomy, and pursuing a goal β€” the ensemble would be treated as an AI agent.

Orchestrated multi-agent systems

For a multi-agent system to fit Article 2(5), the elements of the definition would need to be present together:

  1. Coordination: the systems function as a unified set toward a common objective β€” for example via a central orchestrator, shared memory, or an inter-agent communication protocol.
  2. Perception and action: the set perceives its environment (inputs, queries, system state) and acts upon it (outputs, executing code, modifying data).
  3. Autonomy: the set operates "with a degree of autonomy" β€” making decisions or taking actions without continuous, step-by-step human direction. The degree need not be absolute.
  4. Goal orientation: the set works toward "specific goals."
  5. Adaptation: the set can "adapt to changing inputs and contexts."

A planner agent, a researcher agent and a coding agent that communicate and collaborate on a task would, as a coordinated set, fit the definition.

Edge case: loosely coupled tools vs. coordinated sets

The definition draws a line between coordinated sets and loosely coupled tools.

Loosely coupled tools: Several AI systems that are technically separate and do not coordinate toward a shared goal would not, together, be an AI agent under Article 2(5). For example, one model for email summarisation and a wholly separate model for sentiment analysis remain distinct AI systems; they neither perceive a shared environment in a coordinated way nor act together toward one goal.

Coordinated sets: If those models are integrated so that the sentiment output drives the summarisation strategy, under a unified orchestration layer managing communications autonomously, they could cross into being a "coordinated set." The differentiators are orchestration, a shared goal, and autonomy/adaptation at the level of the set β€” not mere sequential chaining without a feedback loop.

Implications for sovereignty and assurance levels

Why does this matter? Treating a coordinated set as one AI agent would prevent complex architectures from sidestepping requirements by fragmenting their components. Where an AI agent is delivered as, or via, a cloud computing service that a public sector body or Union entity procures, that service could need to meet a Union assurance level under the framework established in Article 16. The sovereignty assessment would attach to the service delivering the set β€” so if part of the system relies on infrastructure or data flows that do not meet the required level, the service could fail to qualify. (CADA’s assurance levels apply to cloud computing services and their providers; they are not a standalone certification of an AI model.)

What this means for you

For CTOs and architects, this definition would change how you document and classify AI infrastructure.

  1. Audit your orchestration layers: If a central coordinator or shared context lets multiple models act autonomously toward a common goal, treat the ensemble as an AI agent β€” not as isolated tools.
  2. Sovereignty by design: When building multi-agent systems for public-sector clients, ensure the cloud service delivering every component can meet the required Union assurance level; a weak link in residency or control could undermine the whole offering’s recognition.
  3. Documentation for audits: Where higher assurance levels are sought, providers face independent third-party audits under Article 20. Your documentation should make the coordination, autonomy and goal-orientation legible, and show how the set adapts to changing contexts.
  4. Risk assessment: Public sector bodies would conduct risk assessments under Article 29 to set the appropriate assurance level. If your multi-agent system supports public-order functions, the service delivering the whole coordinated set would be assessed against the relevant sovereignty criteria.

Common misconceptions

Misconception 1: "A multi-agent system is just several AI systems, so it can’t be one agent." Reality: Article 2(5) expressly includes "a coordinated set of AI systems." The focus is coordinated, autonomous behaviour, not the number of models.

Misconception 2: "If the agents run on different servers, they aren’t coordinated." Reality: Physical location is not the test. Coordination is about logical interaction, shared goals and adaptive behaviour; distributed agents acting together autonomously can still be a "coordinated set."

Misconception 3: "Any tools used together are an AI agent." Reality: Mere co-existence or sequential chaining, without a unified, autonomous, goal-oriented coordination, would not meet the definition. The autonomy and adaptation must exist at the level of the set.

Related

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