Anthropic introduced Managed Agents, a new architecture for AI agents

Anthropic has announced Managed Agents, a service on the Claude platform for executing long-term tasks using AI agents. Its core idea is that agent architecture should remain resilient to the rapid evolution of models, which causes previous constraints to quickly become obsolete.

Traditional agent management systems embed assumptions about model capabilities that eventually lose relevance. For example, the Claude Sonnet 4.5 model used to prematurely terminate tasks due to context limits, but this issue disappeared with the newer Claude Opus 4.5, rendering additional mechanisms unnecessary.

In Managed Agents, the company has divided the system into independent components: the "brain" (the model and control logic), the "hands" (tools and execution environments), and the "session" (an action log). This allows individual parts of the system to be updated separately, making the architecture more flexible—similar to operating systems, where abstractions remain stable despite changes at the infrastructure level.

Abandoning a single container resolved several issues. Previously, a failure would result in session loss and complicate debugging, while also limiting integration with client infrastructure. Now, containers are interchangeable: they can be automatically recreated, and errors are handled as tool failures.

Security has also been strengthened. Previously, generated code was executed in an environment with access to tokens; now, access is isolated and stored outside the sandbox, for example in secure storage, and calls to external services pass through a proxy.

Context handling has been redesigned separately. Instead of rigidly limiting the model's context window, the entire context is preserved in the session log and can be requested as needed. This reduces the risk of data loss in long-running tasks.

According to Anthropic, the new architecture has reduced time to first response by approximately 60% on average and by more than 90% in complex scenarios. Additionally, the system scales better and supports working with multiple tools simultaneously. The company refers to Managed Agents as a "meta-harness"—a universal system designed for the future development of AI. It focuses on modularity and stable interfaces to ensure agents operate reliably even as models become more complex.

Source: anthropic.com