What does self-hosting Claude Managed Agents change for AI sovereignty?

On 12 November 2026, Anthropic opened a self-hosted version of Claude Managed Agents in private beta. The orchestration server, its database and the workers that run tools operate in the customer organization’s own cloud or on premises, and only the requests to the model leave it. Access is requested through the Anthropic account team.

AI agent sovereignty is an organization’s ability to decide, layer by layer, where the model, the orchestration and the tools run, who holds the credentials and session histories, and what crosses its perimeter. It is assessed on the agent’s full architecture, beyond the choice of model provider alone.

What Anthropic announced on 12 November

According to Anthropic’s product article, the harness, sandboxes, credential management and session history can run in your own cloud or on premises. The organization operates the Managed Agents server, its database and the workers where agents run shell and file tools. Secrets can stay in its own secrets manager, and a sandbox’s network limits are set in its own firewall.

Calls to the model go to the destination you choose: Anthropic’s Claude API, Amazon Bedrock, Google Agent Platform, Microsoft Foundry, or an internal gateway. If your organization has zero data retention with Anthropic, it applies to requests sent to Claude, as for any Messages API call.

What the announcement leaves out

The offer is a private beta. No general availability date and no pricing have been shared, and beta terms can change before any wider release. Operations also become the organization’s job: server availability, database backups, updates, worker monitoring. A hurried reader will remember “everything in-house”, while requests to the model still leave, by the route you choose.

Why is the AI sovereignty debate looking at the wrong layer?

The public debate on AI sovereignty centers on the model: the nationality of its publisher, where it was trained, the jurisdiction that governs it. That question matters for the data sent to inference, such as customer file documents. It ignores everything around the model call, which in an agent makes up most of the risk surface.

A production agent is a system: a model that reasons, an orchestration that chains steps and keeps memory, and tools and data it can reach. Each layer has its own possible owner, jurisdiction and failure mode. Treating them separately answers a question that “sovereign or not” never settles: what must stay with us, and why?

The model: where requests go

The first layer is inference. The choice is about the route of requests: straight to the publisher, through a cloud provider already under contract, or through an internal gateway that logs and filters. This route is decided independently of where the orchestration runs, a choice that self-hosted orchestration makes explicit. An internal gateway, for instance, applies the same control to every call, whichever agent issues it. The model remains an external dependency you must be able to replace, a subject covered in our article on model dependency and architecture.

The orchestration: sessions, credentials, history

The second layer is the one the announcement moves. The orchestration holds the agent’s memory, its session history and the credentials it uses: that is where the traces of what it read, decided and executed live. For a data protection officer or a security lead, that history is often more sensitive than the original prompt, because it aggregates data from several sources and intermediate decisions. Keeping it in your own database simplifies retention policy, audit access and incident response.

Data and tools: connectors, secrets, network

The last layer is the most neglected. An agent acts through its tools: connectors to business systems, MCP servers, file access, code execution. An agent’s real sovereignty depends on the rights it is given over those tools, far more than on the model’s nationality. An agent that can write to an ERP with an over-broad account is a risk whether it runs on a public cloud or an internal server. Secrets held in a controlled manager and network egress limited to an allow-list of hosts reduce that risk without removing it. Our analysis of MCP server risks in production details the control points.

For a CIO in a regulated sector, the grid has a practical benefit. It turns a global objection, often phrased as “we cannot use this tool”, into a series of local decisions. One layer may be acceptable as it stands, another may require internal hosting, a third a gateway. The sovereignty objection is then handled layer by layer, with different evidence to provide each time, rather than through a blanket refusal.

How do you apply the layer-by-layer grid to a real agent?

Take an illustrative, generic example: an agent that prepares claim files for a customer service team, reading attachments and proposing a reply that an adviser validates. The example illustrates a method and describes no real deployment.

A decision table for each layer

For each layer, the same set of questions frames the decision:

In the example, the attachments contain personal data. The organization may decide that inference goes through a cloud provider already covered by a data processing agreement, that the history stays in its own database, and that the agent only has read access to the customer file, with writing left to the adviser. Each choice is justified separately, which makes the decision defensible before a risk committee.

What in-house operations require

Hosting the orchestration moves the workload to the platform team. It takes on availability, backups, version upgrades and monitoring, tasks the hosted offer used to carry. An organization without a mature platform team may gain more by keeping the orchestration hosted and hardening the other two layers. The trade-off depends on data sensitivity, operating capacity and the requirements of the sector regulator.

Our trust zones method helps classify agents before choosing where they run: a read-only agent does not call for the same perimeter as an agent that executes payments. Classification comes before architecture, and it is what prevents over-sizing the hosting of low-risk agents.

How do sovereignty, resilience and the AI Act connect?

The hosting choice touches data protection, business continuity and proof of compliance, which are not settled with the same lever.

Resilience: what you can switch over

A supplier outage stops the agents that depend on it. Self-hosting the orchestration does not protect against a model outage, since requests still go there. On the other hand, being able to change the inference route without rewriting the orchestration improves the ability to switch. That means testing the switch before you need it, with a second inference path already under contract. Our article on the model provider turned integrator details the questions to put to suppliers.

AI Act: traceability and human oversight

The AI Act expects high-risk systems to provide traceability, effective human oversight and technical documentation. Under the Digital Omnibus package (Regulation (EU) 2026/1744), the obligations for stand-alone high-risk systems in Annex III are postponed to 2 December 2027. That postponement leaves time to build traceability, provided you decide on it now. A controlled session history with usable logs is the raw material of that proof.

Human oversight weighs just as heavily on the architecture: if an adviser validates an agent’s actions, the orchestration must record who validated what, and when. That detail belongs to the orchestration layer, which the model never sees. Our guide to AI Act compliance for enterprises describes the framework, and the LOOP™ governance page the way to organize it.

Which questions should you ask before requesting beta access?

A private beta suits organizations that already know what they want to test. Before approaching the account team, a few questions prevent a premature request:

The last one is the question people forget. An architecture that depends on a beta with no general availability date must plan its exit, for instance by keeping agent definitions in a format you know how to port to the hosted offer. Your contracts and your legal teams set the exact commitments.

How do you decide without waiting for general availability?

A large part of the work does not depend on the beta. Mapping existing agents, classifying their data and rights, naming an owner for each layer, defining the logs you expect: all of it can start now, with the hosted offer as with self-hosting. When access opens, the decision will only be about where to place the layers.

A good maturity indicator is how quickly you can answer a risk committee’s question: for a given agent, where is the data, who holds the credentials, what history exists? If the answer takes three teams and a week of digging, governance comes before hosting on the list of workstreams. This mapping is useful whatever the final hosting choice, and it makes discussions with a supplier or a regulator more factual.

Our reading: an agent’s sovereignty is managed through a grid of decisions kept up over time, and a label granted once cannot replace it. The AI maturity assessment shows in a few minutes where your governance stands. The enterprise AI governance page details the method, and the enterprise AI agents page the move to production. Koneetiv, an official Anthropic partner, supports this framing; the article on securing Claude agents in production gives the first reflexes.