Anthropic explains Managed Agents architecture for long-running Claude workloads

On April 8, 2026, Anthropic pointed to a new Engineering Blog post in an X post about Managed Agents, its hosted service for long-running agent work on the Claude Platform. Anthropic frames the core problem as building a system for “programs as yet unthought of” and says the service was designed around interfaces that can survive changes in models and harnesses. Instead of treating the agent as one tightly coupled container, Anthropic says Managed Agents virtualizes three pieces separately: the session, the harness, and the sandbox.

The architecture matters because it changes failure handling, deployment flexibility, and security. Anthropic says it decoupled the “brain” from the “hands” and from the persistent session log, letting harnesses restart from a durable event stream and letting execution sandboxes fail independently. In the post, the company says this design reduced p50 time-to-first-token by roughly 60% and p95 by more than 90%. Anthropic also says the change helps customers connect Claude to resources inside their own VPCs without forcing Anthropic’s orchestration assumptions into the customer environment.

Why the separation matters

The blog also argues that the redesign improves the security boundary for agent systems. Anthropic says untrusted code no longer runs in the same place as long-lived credentials, and that OAuth tokens for MCP tools can stay in a separate vault while Git credentials are wired in during sandbox initialization. That matters because long-horizon agents are increasingly judged less by raw benchmark scores than by whether they can recover from failure, resume work, and handle sensitive infrastructure safely. Managed Agents looks like Anthropic’s answer to that operational layer, not just another packaging change around Claude.