MAQPNA

Learn

The EU AI Act and AI agents: logging, oversight and deployer duties.

What Article 12, Article 14 and Article 26 ask of an agent in production, and what evidence shows you did it.

Updated 31 October 2026. This page is not legal advice.

Does the EU AI Act apply to AI agents?

Yes, but not because they are agents. The EU AI Act (Regulation (EU) 2024/1689) regulates AI systems by risk and use. An agent carries the heaviest duties when it is a high-risk AI system: used for an Annex III purpose such as recruitment, credit scoring or access to essential services, or built into a product listed in Annex I. Then providers must design it for logging (Article 12) and human oversight (Article 14), and deployers must use, watch and keep its logs (Article 26).

Most agents that write code or answer tickets are not high-risk. They may still meet the transparency duties of Article 50 (telling people they are dealing with AI) and the AI literacy provisions of Article 4, which the Digital Omnibus also revised. The model underneath may fall under the general-purpose AI rules, which bind the model provider, not you. Readable article texts are on artificialintelligenceact.eu.

When do the obligations apply after the Digital Omnibus?

The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It moved the high-risk dates. It is adopted law, not a proposal; the dates below follow law-firm summaries such as Orrick's, so check the consolidated text for your case.

DateWhat applies
1 August 2024The AI Act enters into force
2 February 2025Prohibited practices and AI literacy (Article 4)
2 August 2025General-purpose AI model obligations, governance and penalties
2 August 2026Transparency obligations (Article 50)
2 December 2026Marking of AI-generated content (Article 50(2)) for systems already on the market before 2 August 2026
2 December 2027High-risk systems under Article 6(2) and Annex III, previously 2 August 2026. Systems already on the market are covered only if substantially modified afterwards
2 August 2028High-risk systems in Annex I regulated products
2 August 2030High-risk systems used by public authorities

What does Article 12 logging require for an agent?

Article 12 says a high-risk AI system must technically allow the automatic recording of events (logs) over its lifetime. The logs must make its functioning traceable: they must help identify situations that may present a risk or a substantial modification, support the provider's post-market monitoring (Article 72) and support the deployer's monitoring under Article 26(5). Providers keep the logs under their control for at least six months (Article 19); deployers do the same under Article 26(6).

For an agent, the events that matter are its actions: each tool call, model call and call to another agent, who it acted for and what was decided. Application logs can be edited by whoever runs the application. A log an auditor can rely on should show that nothing was changed, removed or reordered. That is the case for a tamper-evident audit trail.

What does Article 14 human oversight mean for an agent?

Article 14 requires high-risk systems to be designed so that people can oversee them effectively while in use, in proportion to the risk and the system's level of autonomy. The people overseeing must be able to monitor it, interpret its outputs, decide not to use it or to disregard, override or reverse an output, and interrupt it through a "stop" button or a similar procedure that brings it to a safe halt.

An agent acts faster than a person reads. Oversight therefore has to sit where actions happen: a rule that holds a risky call for a human, a kill switch that stops an agent mid-run, and a timeline that shows what it did. Article 14(5)'s two-person check applies only to remote biometric identification, but four-eyes approval follows the same idea for high-impact actions.

What must deployers of high-risk agents do under Article 26?

Article 26 is the deployer's list. Use the system as its instructions say. Give oversight to people with the competence, training and authority to do it. Make sure input data you control is relevant. Monitor the system, and if it presents a risk, tell the provider and suspend use. Keep the logs under your control for at least six months. Inform workers before using it at work, and inform people when a high-risk system makes or helps make decisions about them.

Providers carry two further duties that touch agents in production. Article 72 requires post-market monitoring, which draws on the logs. Article 73 requires serious incidents to be reported to the market surveillance authority within strict deadlines. Deployers who find one must tell the provider first.

Which evidence answers which obligation?

"Supports evidence for" means MAQPNA produces records you can use. It never means the obligation is met.

ObligationWhat evidenceMAQPNA feature
Art. 12 automatic loggingA complete, unaltered record of each action, who it was for and the decisionAudit ledger: SHA-256 hash chain, signed checkpoints, offline maqpna audit verify; argument hashes, not raw arguments
Art. 19 and 26(6) six-month retentionProof logs are kept long enough and cannot be deleted earlyThe eu-ai-act compliance profile refuses retention below 183 days; write-once storage with Object Lock; a metric when retention falls below the minimum
Art. 14(4)(d) override or reverseRecords of held actions and the human decisionApprovals with four-eyes quorum and no self-approval; approver identity in the ledger
Art. 14(4)(e) stopA tested way to halt the system, and a record of each useKill switch by agent, session, user or installation; each revocation recorded
Art. 14(4)(a) and 26(5) monitoringA view of what the agent is doing and why calls were deniedSession timeline, alerts, OpenTelemetry GenAI traces, SIEM forwarding
Art. 72 and 73 monitoring and incidentsA verified timeline of what happened in a windowLedger export with verified hash chain, denials, approvals and revocations; the provider still writes and files the report
Auditor review for a system and periodOne signed bundle tying the above togetherEvidence pack for eu-ai-act, verifiable offline
Art. 9, 10, 11 risk management, data governance, technical documentationProvider documents and processesNot covered by MAQPNA (customer process)
Art. 26(7) and 26(11) informing workers and affected people; Art. 27 fundamental rights impact assessmentNotices, consultations and assessmentsNot covered by MAQPNA (customer process)
Art. 4 AI literacy; Art. 50 transparency noticesTraining records; disclosure in your productNot covered by MAQPNA (customer process)

How do DORA and NIS2 bear on AI agents?

DORA (Regulation (EU) 2022/2554) has applied to EU financial entities since 17 January 2025. An agent that calls external MCP servers and model APIs uses ICT third-party services, which belong in the register of information required by Article 28(3). Major ICT-related incidents must be reported under Article 19. MAQPNA's maqpna evidence register builds a third-party register of every MCP server, model route and agent peer that agents actually used, from the ledger. It is an input to your register, not the register itself.

NIS2 (Directive (EU) 2022/2555) requires essential and important entities to take risk-management measures under Article 21, including incident handling, supply-chain security, access control and asset management. Article 23 sets reporting stages: an early warning within 24 hours, a notification within 72 hours and a final report within one month. Agents with tool access are part of that attack surface. Short-lived identities, default-deny egress, tool pinning and a verifiable ledger support those measures; national transposition and the reporting itself remain yours.

Where MAQPNA fits

Controls and evidence, not a compliance badge.

MAQPNA is a runtime that sits around your agents. Every tool, model and agent-to-agent call goes through one gateway that can log, hold and stop it. That gives you the raw material for Articles 12, 14 and 26. It does not classify your system, write your technical documentation or file reports, and Kosa8 Labs holds no certifications. The evidence pack mapping is marked draft until counsel and audit partners review it.

Related reading: AI agent governance, the OWASP agentic top 10 mapping, ISO 42001 and the NIST AI RMF, sovereign and air-gapped governance, the trust centre and solutions by sector.

FAQ

Common questions

Is an AI agent a high-risk AI system?

Only if its use is high-risk: an Annex III purpose such as recruitment or credit scoring, or a safety component of an Annex I product. A coding or support agent is usually not high-risk, though Article 50 transparency may still apply.

When do the high-risk obligations apply?

After the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), Annex III high-risk obligations apply from 2 December 2027 and Annex I from 2 August 2028.

How long must agent logs be kept under the EU AI Act?

At least six months for both providers (Article 19) and deployers (Article 26(6)), unless other EU or national law sets a different period. Financial institutions keep them under their financial-services rules.

Does MAQPNA make us EU AI Act compliant?

No product can. MAQPNA supplies runtime controls and signed evidence for logging, oversight and monitoring; risk management, documentation, notices and reporting stay with you.

Keep the records before you need them.

Run agents through one gateway that logs, holds and stops, then build a signed EU AI Act evidence pack from the ledger.

More install options
curl -fsSL https://maqpna.com/install.sh | sh