EU clarifies agents under AI Act

Europe’s artificial intelligence law now reaches the software “agents” that can plan and carry out multi-step tasks, even though the law never names them directly. (digital-strategy.ec.europa.eu) The European Commission’s guidance says the Artificial Intelligence Act applies by function, not by product label: if a system uses a general-purpose model or falls into a high-risk use, the rules attach to that system. The law itself starts applying in full on August 2, 2026, after earlier provisions for banned uses and general-purpose models took effect in 2025. (digital-strategy.ec.europa.eu) (artificialintelligenceact.eu) An artificial intelligence agent is software built to take actions with less human prompting, like managing an inbox or booking travel. The Future Society said in a June 4, 2025 analysis that the European Union’s law was not drafted around agents, but that its rules still govern them through the categories for general-purpose models and high-risk systems. (thefuturesociety.org) That means compliance turns on what the agent does, and who built or deploys it, rather than on whether a company markets it as an “agent.” Under the Act’s risk-based structure, some uses are banned outright, while higher-risk systems face documentation, oversight and conformity duties. (digital-strategy.ec.europa.eu) (eur-lex.europa.eu) The timing matters because the Commission says its enforcement powers for general-purpose model obligations apply from August 2, 2026. Providers of models with systemic risk already have legal duties to notify the European Union’s Artificial Intelligence Office and work toward compliance. (digital-strategy.ec.europa.eu) The law also reaches companies outside Europe if their systems are placed on the European Union market or used in ways covered by the Act inside the bloc. That gives the guidance weight for United States and other foreign firms selling agentic tools into Europe. (digital-strategy.ec.europa.eu) (barradvisory.com) A separate UNESCO-backed snapshot published April 14, 2026 suggests many companies are still not ready for that kind of governance. The report covered 2,972 companies across 11 sectors and found 43.7% publicly said they had an artificial intelligence strategy or guidelines, but only 13% publicly claimed adherence to a formal governance framework. (dig.watch) The same report said 72% of companies did not report any artificial-intelligence impact assessment, 40% reported board- or committee-level oversight, and 12.4% reported a policy requiring human oversight of artificial-intelligence systems. On labor, 31% said they had training programs, but only 12% described structured training with comprehensive coverage. (dig.watch) European regulators are not creating a separate “agent law” here. They are signaling that when software can act more independently, the existing Artificial Intelligence Act follows the model, the use case and the company in the chain that controls the risk. (digital-strategy.ec.europa.eu) (thefuturesociety.org)