With “Claude Managed Agents,” Anthropic has introduced a new infrastructure solution designed to make it easier for developers to operate AI agents. The service, available immediately as a public beta, takes over complex background tasks such as session management and security sandboxes in order to shorten the path from prototype to market readiness.
Until now, many AI agent projects have failed not because of the intelligence of the underlying language models, but because of the technical complexity of their environment. Anyone wanting to build autonomous agents had to grapple intensively with server infrastructure, securing code execution, and managing permissions. This is where Anthropic steps in, now offering these components “as a service.”
A Focus on Infrastructure and Security
The core promise of Claude Managed Agents lies in taking the burden off development teams. Instead of building their own solutions for executing tools, developers get an integrated environment from Anthropic. The most important pillars of the new offering include:
- Secure sandbox environments: Agents can execute code in isolated areas, which minimizes the risk to the rest of the system.
- Long-running sessions: While conventional API requests are often short-lived, Managed Agents support sessions that remain stable over several hours.
- Integrated tracing: For debugging and optimization, the platform provides tools that allow the AI’s decision steps and tool calls to be traced in detail.
In addition, Anthropic is experimenting with coordinating multiple agents (“multi-agent coordination”). However, this feature is currently still in an early research phase (research preview) and is not yet intended for full production use.
Pricing Model and Availability
Billing for Managed Agents follows a hybrid model. In addition to the usual costs for the volume of data processed (tokens), Anthropic charges an extra fee for providing the infrastructure. This amounts to 0.08 US dollars per active session hour. With this, the company positions itself in a price range that is especially attractive for companies that prioritize scalability over building their own infrastructure.
The service is accessible immediately as part of a public beta via the Claude platform for registered developers.
Background: Key Terms Explained
- Sandbox: An isolated virtual area in which programs (or AI agents) can be executed without having access to the main system. This serves to ensure security.
- Token: The basic unit of computation for language models. One token corresponds to roughly part of a word. The costs for AI models are usually calculated per one million tokens processed.
- Tracing: A method for logging program flows. In the AI context, it helps to understand why an agent made a particular decision or which tool it used.
- Public beta: A testing phase in which software is already publicly accessible but may still contain bugs.
Sources: Anthropic product announcement, Claude Platform technical documentation.
