Part 32 · One Agent, Five Platforms
Regulated? Then where the agent and its data run matters as much as what it does. An invoice agent reads supplier bank details, approval history and amounts. A security or data protection review will ask three questions before it asks anything about prompts:
- Runtime: where does the agent itself run, and can it run in our own infrastructure?
- Model: where are prompts and completions processed?
- Logs: where do traces, conversation history and audit records end up?
Part 30 flagged residency as the control with the most gaps. This part answers the three questions for each platform, from vendor documentation I read on 20 November 2026. Like the builds, it's documentation-only, and the gaps are stated rather than filled.
MuleSoft
- Runtime. Agent networks deploy to a CloudHub 2.0 shared or private space, or a Runtime Fabric target. Runtime Fabric runs on infrastructure you manage, so it's the customer-hosted option (Part 11). The control planes overview lists agent network projects and the LLM gateway as available on the EU control plane. Since 16 November 2026, the release notes also list Hyperforce as a target for version 2.0 agent networks.
- Model. The broker calls an LLM provider you configure, through the gateway. I found no statement about where those calls are processed. That depends on the provider and region you choose.
- Logs. Traces and logs go to Anypoint Monitoring, and the EU control plane lists log search, raw log data and a telemetry exporter. Retention and storage location weren't on the pages I read.
Boomi
- Runtime. The runtime clouds page lists four Boomi-managed clouds: US East, UK, Japan and Australia. There's still no EU cloud; the UK one is outside the EU. A self-hosted Flex runtime "runs in your own infrastructure" but is in early access, as it was in Part 15.
- Model. The Boomi AI FAQ says agents use models through Amazon Bedrock, and that each runtime runs the model, prompts and responses "in the selected region". You can also connect your own Anthropic, OpenAI or Google account, in which case prompts go to that provider. On Flex, "only the model prompt leaves your network".
- Logs. Audit trace logs are in Agent Control Tower, without a stated retention period. Chat sessions are kept until a user deletes them. Two exceptions matter for residency: glossary and knowledge base content is stored in the platform's region, so a North America customer running agents in the UK, Japan or Australia retrieves it from the US.
SAP
- Runtime. The Business AI reference architecture says the platform is "SAP-managed by default", with SAP AI Core running on AWS, Azure and Google Cloud. SAP Learning's data residency lesson describes locally hosted AI, SAP Sovereign Cloud and an EU AI Cloud with "AI infrastructure hosted within Europe". I found no on-premises option for Joule agents, and no availability dates for those offerings.
- Model. Models come through the generative AI hub, from SAP and partners such as Anthropic, OpenAI, Mistral and Cohere. The lesson says processing "may occur on SAP-managed infrastructure or approved external model providers, depending on the use case and configuration". I found no way to pin Joule's model calls to a region.
- Logs. Not covered in what I read, as in Part 18. The architecture mentions agent session telemetry going to SAP Cloud ALM, but not where it's stored or for how long.
Azure
- Runtime. Foundry Agent Service endpoints are regional, and the FAQ says "data is stored in the same region as the endpoint". In the standard setup, threads, files and vector stores go to your own Cosmos DB, Storage and AI Search resources. Agents can run in your virtual network, on a delegated subnet in the same region. There's no on-premises option on these pages.
-
Model. This is where Azure is clearest. The deployment types page defines three levels:
- Global: may be processed in any Azure region.
- Data Zone: processed only within the US, EU or APAC zone. The EU zone is the EU Data Boundary.
- Standard (regional): processed within the geography you choose. In every case, "data stored at rest remains in the designated Azure geography". So the deployment type you pick is a residency decision, not just a pricing one.
- Logs. Traces go to an Application Insights resource in your own subscription (Part 22).
AWS
- Runtime. AgentCore Runtime, Gateway, Identity and Observability are available in seven European regions, according to the regions page: Frankfurt, Ireland, London, Milan, Paris, Spain and Stockholm. Milan and Spain don't have Memory or Runtime Instances yet. The runtime can attach to your VPC (Part 25). There's no on-premises option.
- Model. Bedrock's geographic cross-Region inference keeps EU requests within EU regions. But prompts and results "might move outside of your source Region", and any abuse-detection copies are stored in the destination region. On the AgentCore side, the cross-Region inference page says Memory, Policy and Evaluations use it. From the EU regions it stays in Europe. From some other source regions, Policy and Evaluations use global routing that can leave the geography.
- Logs. Observability writes to CloudWatch in your account (Part 25). The cross-Region page warns that CloudWatch and CloudTrail logs "won't specify the AWS Region in which data inference occurs".
What the builds could and could not show
- Could: where each vendor says the runtime runs, which regions exist, and which platforms offer a customer-hosted runtime. MuleSoft has Runtime Fabric, and Boomi has Flex in early access.
- Could not: where model calls actually land for MuleSoft and SAP, and log retention on most of them. Those need the vendor's trust documentation or contract terms, and a hands-on build that checks the traces.
- For an EU-only requirement: Azure's Data Zone and AWS's EU regions are the most explicit in the docs. MuleSoft can keep the runtime and control plane in the EU but leaves model location to your provider. Boomi has no EU-managed cloud. SAP describes an EU AI Cloud without dates I could find.
The policy from Part 30 still holds: classify the data first, then pick the runtime, model deployment and log store that keep each class where it's allowed.
Where must your agent data stay?
This series describes a reference build on a fictional company. This part is based only on vendor documentation read on 20 November 2026 and the documentation-only builds in Parts 11 to 30; no hands-on build has been run yet. It's not legal or compliance advice. Views are my own.

Top comments (0)