The bottom line: Anthropic decouples the components of its Managed Agents: Session, Harness, and Sandbox now run independently, making systems more reliable, easier to debug, and more future-proof—similar to how operating systems use hardware virtualization to enable programs that don’t yet exist.
Anthropic introduces Managed Agents—a hosted service that operates long-running AI agents over standardized interfaces. The concept decouples an agent’s core components from one another to make systems more future-proof and maintainable.
Managed Agents follow an established pattern from computer science: virtualizing hardware components into abstract interfaces that have longer lifespans than individual implementations. Just as operating systems virtualized memory and processes decades ago, Managed Agents now virtualize the core elements of an AI agent.
The system consists of three decoupled components: a Session (the immutable protocol of all events), a Harness (the loop that calls Claude and forwards tool calls), and a Sandbox (the execution environment for code and files).
Anthropic initially bundled all components into a single container. However, this led to classic infrastructure problems: the container became a “pet” instead of exchangeable “cattle.” When containers failed, the session was lost. Debugging was difficult because errors in the Harness, network problems, or container crashes all looked identical. Additionally, the Harness couldn’t easily be connected to external systems like Virtual Private Clouds.
By decoupling the components, these problems are solved: the Harness can be updated independently without losing sessions. Each component can be debugged and scaled individually. This also enables model improvements—such as fixing “Context Anxiety” in Claude Sonnet 4.5—to be implemented without changes to the entire system.
Source: www.anthropic.com