The EU AI Act high-risk system requirements take effect on August 2, 2026. If you are deploying AI agents in the EU, or deploying agents that process data from EU residents, you have roughly five months to demonstrate compliance. This is not a suggestion. The penalties are up to 35 million EUR or 7% of global annual turnover, whichever is higher.
Most teams building with AI agents have not mapped the Act's requirements to their actual infrastructure. Here is what the key articles demand and what they mean in practice.
The Act requires that high-risk AI systems "automatically record events" including the period of use, the reference database, input data, and the identification of natural persons involved in verification. For AI agents, this means every action decision must be logged with enough context to reconstruct what happened.
Generic application logs do not satisfy this. You need structured records that capture the intent (what the agent wanted to do), the policy version that evaluated it, the rules that matched, the decision, and the outcome. This is exactly what Authensor's receipt system provides. Every authorization decision produces an immutable, hash-chained receipt with all of these fields. Receipts are queryable, exportable, and tamper-evident.
The Act requires that high-risk systems can be "effectively overseen by natural persons" and that humans can "intervene on the operation of the high-risk AI system or interrupt the system." For agents, this translates directly to approval workflows.
You need the ability to flag certain actions for human review before they execute. Not after. The Authensor policy engine supports REVIEW decisions that route intents to human approvers. The agent pauses, the human reviews the intent with full context, and approves or rejects it. The decision is recorded in the receipt chain. This is Article 14 compliance in practice.
The Act requires a risk management system that identifies, analyzes, and mitigates risks throughout the lifecycle. For agents, this means ongoing behavioral monitoring and adversarial testing.
Sentinel, Authensor's behavioral monitoring engine, tracks per-agent metrics using statistical methods (EWMA for trend detection, CUSUM for shift detection). It flags anomalies in deny rates, latency, action volume, and multi-agent delegation depth. This is continuous risk monitoring. On the testing side, Authensor's RedTeam toolkit provides structured adversarial evaluation to identify vulnerabilities before deployment.
The Act requires that high-risk systems are designed to be "sufficiently transparent to enable deployers to interpret the system's output and use it appropriately." For agents, this means decision explainability.
Every Authensor policy decision includes an explanation: which rules matched, why the action was allowed or denied, what conditions were evaluated. This is not a black box. An auditor can trace any agent action back to the specific policy rule that governed it, the version of the policy that was active, and the exact input that triggered the evaluation.
Article 12 also requires that logs be retained and accessible. If your compliance infrastructure runs on a third-party SaaS, you are adding a dependency to your compliance chain. If that service goes down, changes its API, or gets acquired, your compliance posture is at risk.
Authensor is open source and self-hostable. You run it on your infrastructure. You own the data, the receipts, the policies, and the audit trail. No vendor lock-in. No data leaving your perimeter unless you choose to send it somewhere.
The deadline is August 2, 2026. Start by mapping which of your agent deployments qualify as high-risk under the Act, then build the logging, oversight, monitoring, and transparency infrastructure to cover them. Authensor provides all four layers. The code is public. Start with npx create-authensor and read the compliance mapping in the docs.