The EU AI Act doesn't have a separate category for AI agents. Instead, it regulates AI systems based on how they're used — and an agent's obligations depend on what it does. This overview explains how the EU AI Act relates to AI agents, which requirements matter most in practice, and the current timeline. It's general information, not legal advice. Your legal team should assess how the Act applies to your systems.
How the EU AI Act classifies systems
The Act takes a risk-based approach:
- Prohibited practices — a short list of uses that are banned outright.
- High-risk systems — uses listed in the Act (for example in employment, access to essential services, or certain safety components) that carry the heaviest obligations.
- Transparency obligations — for example, telling people they're interacting with an AI system, and labelling certain generated content.
- General-purpose AI models — obligations on model providers.
- Everything else — largely voluntary codes and good practice.
An agent that answers internal IT questions is unlikely to be high-risk. An agent involved in decisions about hiring or creditworthiness may be. The use case decides, not the technology.
Requirements that matter most for agents
For high-risk systems, several obligations map directly onto how agents are built and run:
Human oversight (Article 14)
High-risk systems must be designed so that people can oversee them effectively — understanding their operation and being able to intervene. For agents, that usually means defined approval points for consequential actions and the ability to stop the agent. See human on the loop vs human in the loop.
Record-keeping (Article 12)
High-risk systems must allow automatic logging of events over their lifetime. Under Article 26, deployers must keep the logs under their control for at least six months, unless other law requires otherwise. For agents, a decision-level AI agent audit trail is the natural foundation.
Deployer obligations (Article 26)
Organisations using high-risk systems must use them according to instructions, assign oversight to competent people, monitor operation, and report serious incidents.
The timeline after the Digital Omnibus
The Act entered into force in August 2024 and applies in stages. In 2026, the "Digital Omnibus on AI" amended the timeline. As reported at the time of writing:
- Prohibitions and AI literacy — applying since February 2025.
- General-purpose AI model obligations — applying since August 2025.
- Transparency obligations (Article 50) — 2 August 2026.
- High-risk systems listed in Annex III — moved to 2 December 2027.
- High-risk systems in products covered by Annex I — moved to 2 August 2028.
Dates and details can change; check the official text and current guidance.
What to do now
Even if your agents aren't high-risk, the practices the Act describes are good engineering:
- Inventory agents and what each one does — see AI agent risk assessment.
- Define human oversight for consequential actions.
- Log decisions, not just model traces.
- Keep records of who changed agent rules, and why.
These also support AI agent compliance reviews under other frameworks.
Where Gaten fits
Gaten helps you implement the practices behind oversight and record-keeping for agents: approval points for consequential actions, and a decision log with the rule and approver for each action. It doesn't determine your legal obligations or make you compliant on its own. See the audit trail.
Frequently asked questions
Does the EU AI Act apply to AI agents?
It applies to AI systems based on their use. An agent may fall under high-risk, transparency or no specific obligations depending on what it does.
When do the EU AI Act high-risk obligations apply?
Following the 2026 Digital Omnibus amendment, obligations for high-risk systems listed in Annex III were moved to 2 December 2027, and for those in Annex I products to 2 August 2028. Check current official guidance.
How long must logs be kept under the EU AI Act?
For high-risk systems, deployers must keep automatically generated logs under their control for at least six months, unless other applicable law says otherwise.

