Skip to content
Sentynel

#technical5 min readBy Sentynel

Self-hosted agents move the server, not the boundary: map what crosses it

Anthropic now lets companies run its managed agent service on their own infrastructure. Here is how to map what still leaves, and what you now have to run.

Share

On 12 November 2026 Anthropic opened a private beta of self-hosted Claude Managed Agents. In its words, "every part of the service runs in your own cloud or on-premise, including the harness, sandboxes, credentials, and session history," while "model requests go where you choose." That moves the servers inside your perimeter. It does not decide which parts of your operation's data should reach a model, and that decision is a design job, step by step.

Hosting is a location. The boundary is a list.

Moving a service on-premise answers "where does it run?" It does not answer "what leaves?" Anthropic says it plainly: model requests still go to a provider you pick, whether that is Anthropic directly, a cloud provider or your own gateway. Every prompt an agent sends carries whatever the step put into it: a customer's message, an order history, a contract clause.

So the real boundary is not a network diagram. It is a list, per step, of what goes into each model call. Most companies have never written it, because when the whole agent ran at a vendor it seemed pointless. Self-hosting makes the list useful, because now you control most of the path and can actually enforce it.

The four things that now live in your house

According to the announcement, a self-hosted deployment means you operate "the Managed Agents server, its database, and the workers where agents run their shell and file tools." Translate that into operations work:

  1. The harness. The loop that decides what the agent does next. Its version is now your release, so it needs the same change control as any prompt or tool change.
  2. Sandboxes. Where tools execute. You decide what each one can reach: which systems, which folders, which network routes.
  3. Credentials. The keys to your CRM, ERP and messaging channels. They now sit in your own secrets store, which means rotation and revocation are your procedures, not a vendor setting.
  4. Session history. Anthropic notes it "is stored in your own database." That is a full record of what every agent read, said and did. It is the most valuable operational log you will own, and also a store of personal data with its own retention obligations.

None of these runs itself. Each one needs a named owner before the first agent goes live.

Map the boundary step by step

Take one flow, say an incoming customer enquiry that ends in a resolved case, and write one row per step. Five columns are enough:

Column Question
Step What the agent does, in one verb (classify, look up, draft, send)
Reads Which records and fields it needs, by name
Sends to the model The fields that actually go into the prompt
Writes What it changes, and in which system
Owner The person who answers for this step

The gap between "Reads" and "Sends to the model" is where the design happens. A step that classifies an enquiry may need the message text but not the customer's tax ID. A step that drafts a reply may need the order status but not the payment details. Every field you can leave out is a field that never crosses.

One step, mapped at the boundary
  1. 01Read the recordInside: CRM, orders, history
  2. 02Minimize fieldsOnly what the step needs
  3. 03Model callCrosses to the chosen provider
  4. 04Act in a sandboxInside: scoped tools and keys
  5. 05Write the journalInside: session history you own

Review the map whenever a step, a tool or the provider changes.

Five rules for the minimization layer

  • Name fields, never tables. "Order status and delivery date" can be checked. "Order data" cannot.
  • Put the filter in code, not in the prompt. Telling the model to ignore a field does not stop the field from being sent. Strip it before the request is built.
  • Classify once, reuse everywhere. Tag fields as public, internal, personal or restricted in the system of record, and let every step read the tag. A new agent then inherits the rule instead of reinventing it.
  • Route by class, not by agent. If some data may only go to a provider in a given region or through your own gateway, make that a routing rule on the data class. The announcement lists the gateway as an option; the rule for when to use it is yours.
  • Log what was sent, not only what was answered. The session history should let an auditor see the exact fields that left for each call. If it only stores the agent's replies, you cannot prove the boundary held.

The session history is now an operations asset

Most teams treat agent logs as debugging output. Once they live in your own database, they become the case record of the work: who asked, what the agent saw, what it decided, what a person changed.

That is the material operations improvement runs on. In Rodo's service operation, built by Sentynel, the measured result is "82% of live-created enquiries resolved". It could only be stated because every enquiry had a record with a stage. The case record notes it covers 31 August–24 September and describes pipeline records, not all conversations.

Rodo · first measured window
of live-created enquiries resolved

Source: Source: Rodo operational report, 24 September 2026. The resolution cohort covers 31 August–24 September. Figures describe pipeline records, not all conversations.

Decide three things about the history before launch:

  1. Retention. How long each class of data stays, and who can extend it for an open case.
  2. Access. Who may read full sessions, and who sees only the case-level summary.
  3. Use. Which metrics are computed from it, such as time to resolution or the share of cases handed to a person.

What self-hosting does not change

The announcement notes that developers "still build agents with the same core APIs as the hosted Managed Agents service." That is useful: you can design the flow once and choose where it runs. It also means the hard part stays the same in both places. An agent with a vague job, unowned exceptions and no journal is just as weak on your servers as on anyone else's.

The feature is in private beta, requested through Anthropic's account team, so few companies will run it this quarter. The boundary map does not have to wait. It works the same with any provider and any hosting choice, and it is the document a security review will ask for first.

A test for this week

Pick the agent flow that touches the most personal data. Write the five-column table for it. For each step, circle every field in "Sends to the model" that the step could do without. If you circle nothing, look again. If you circle more than you keep, you have found your first redesign.

Sources

  1. https://claude.com/resources/articles/self-hosted-claude-managed-agents
Share
← All sygnals