Agent framework landscape in Europe

Three regulatory instruments reached distinct stages of operational weight between 2024 and 2026. The EU AI Act's General-Purpose AI provisions applied from August 2024. The high-risk provisions under Annex III, which cover agent configurations used in credit scoring, fraud detection, and automated financial advising, do not apply until August 2026, but the conformity assessment preparation cycle for those applications is already underway. GDPR enforcement activity against international data transfers intensified following the Schrems II consequences working through national supervisory authorities. PSD2's open banking mandate matured to the point where API-based payment orchestration became a standard architectural component rather than an advanced experiment [6][7].

Agent frameworks sit at the intersection of all three instruments. A framework that orchestrates multi-step tasks on behalf of a user in a financial services context is simultaneously subject to AI Act risk classification (if its outputs affect credit, insurance, or investment decisions), GDPR transfer obligations (if its inference or memory components process personal data outside the EU), and PSD2 licensing requirements (if it initiates or facilitates payment transactions) [9][12].

Framework deployment preceded the convergence of binding obligations, leaving operators without settled guidance on which configurations trigger which requirements. The consequence is that operators evaluating frameworks now are self-certifying compliance posture against instruments whose application to multi-agent architectures is still being defined. The three operational axes this brief examines are not abstract rating criteria; they are the specific dimensions on which supervisory scrutiny will land when the August 2026 high-risk provisions become enforceable.

Framework selection for European agentic deployments involves three independent compliance dimensions, each of which can disqualify an otherwise capable system on its own terms.

Data residency operates at the infrastructure layer. GDPR Chapter V prohibits personal data transfers to third countries without an adequacy decision, standard contractual clauses with a transfer impact assessment, or a binding corporate rules regime. For agent frameworks, this means that model inference endpoints, memory stores, tool call logs, and orchestration telemetry must each be evaluated as potential personal data flows. A framework that routes these components through non-EU infrastructure (even transiently) creates a transfer liability that persists regardless of the framework's domestic licensing status. Commercial hyperscaler offerings typically provide EU-region deployments, but contractual protections in standard tiers frequently fall short of the GDPR Chapter V standard without negotiated addenda [9][12].

Open-source governance operates at the supply-chain and auditability layer. An open-source framework's licence determines whether derivative works carrying EU-regulated logic (credit scoring, fraud detection, automated advising) can be distributed under proprietary terms, and whether the framework owner can compel downstream operators to disclose modifications. Separately, the structural completeness of the governance artefacts that practitioners attach to agent configurations determines whether an agent system can demonstrate the accountability trail the EU AI Act requires for high-risk applications [10]. Governance file completeness is therefore an operationally testable criterion, not merely a normative aspiration.

EU infrastructure integration operates at the protocol and certification layer. PSD2-compliant Open Banking APIs require Strong Customer Authentication and explicit consent flows that differ structurally from generic API calls [6][7]. Agents that initiate or orchestrate payment actions must implement these flows at the framework level or accept that every deployment team will rebuild them independently, generating inconsistent compliance posture across operators. The same logic applies to identity: agent systems interacting with EU residents on behalf of regulated entities need a mechanism to assert, verify, and delegate identity in terms that eIDAS 2.0 and future EUDI Wallet infrastructure will recognise. No current framework provides for this at the framework level; the gap sits at the framework-to-protocol interface and is not resolved by application-layer workarounds alone [12].

In the near term (zero to twelve months), the absence of framework-level payment integration means that operators building payment-adjacent agents must budget for bespoke connector development, independent SCA implementation, and the compliance review cycle that accompanies each. SCA-compliant flows interacting with SEPA Instant require tested interoperability that framework vendors have not published, and that burden falls on the licensed institution. Procurement teams evaluating commercial frameworks should require contractual commitments on EU-region inference endpoints, auditable transfer impact assessments, and documented AI Act conformity assessment status before signature. Standard enterprise agreements from major hyperscalers do not currently carry these provisions in their base tier [9].

In the medium term (one to three years), the EU AI Act's high-risk classification provisions under Annex III will apply from August 2026 directly to agent frameworks used in credit scoring, fraud detection, and automated financial advising. Operators who adopted frameworks under a compliance-neutral assumption will face conformity assessment requirements against a foundation that was not designed to produce the required technical documentation. Open-source frameworks that permit full auditability of training data provenance, inference logic, and output logging will be structurally better positioned for this assessment than commercial black-box alternatives, provided their governance artefacts satisfy the completeness threshold [10].

Over the longer horizon (three-plus years), regulatory competition between EU-sovereign framework initiatives and incumbent hyperscaler offerings will determine whether the platformisation dynamic identified in payment infrastructure [6] reproduces itself in the agent framework layer. If it does, the operator community will face the same concentration and switching-cost structure that PSD2 created in open banking: technically open by mandate, but commercially captured in practice [7][8].