Managed Agents vs. Self-Hosting
Explore the operational and compliance considerations of running Claude agents using managed hosting versus self-hosting. Understand the components of Anthropic's Managed Agents service, compare it with self-hosted deployments, and learn how to recommend the best hosting option based on a customer’s security, budget, and control needs.
Building an agent is half the work. A customer also needs it to run every day, for many users, with its records kept and its failures noticed. Someone has to operate the program that runs the agent, and that choice shapes cost, compliance, and who gets paged when something breaks.
Claude agents can run in two places. We can run the agent ourselves, on infrastructure the customer or we operate, or we can let Anthropic run it as a hosted service. FDEs make this recommendation early in an engagement, because it affects the customer’s security review, budget, and team. The choice also has to hold up after handoff, when the customer’s own engineers operate the agent without the FDE in the room.
Two places an agent can run
Every agent depends on an agent harness, the program that runs the agent loop. It sends requests to Claude, runs the tools Claude asks for, keeps the conversation state, and decides when the task is done. The question is who operates that harness.
Self-hosting: Our application runs the harness, usually with the Agent SDK, on servers or containers that the customer or we manage. The agent loop runs in our process.
Managed hosting: Anthropic runs the harness on its own infrastructure through Claude Managed Agents. Our application describes the agent, sends it messages, and receives a stream of events.
The difference resembles running our own database server versus ...