Anthropic has launched Claude Managed Agents in public beta, offering a containerized agent solution that runs natively in Claude’s cloud. This new feature includes a credentials vault for securely storing OAuth and bearer tokens, environments for configuring isolated container sessions with customizable network access and package settings, and sessions that can be created manually or automatically by agents. Users can also mount files or GitHub repositories to sessions, allowing for granular control over the agent’s working environment. The interface supports both YAML and JSON configurations, with Claude assisting in building these setups through templates and quick start guides.
The key distinction between managed agents and traditional skills lies in their use cases and operational scope. Skills are ideal for personal workflows running on a user’s machine, providing access to tools and MCP servers within that context. In contrast, managed agents are designed for multi-user scenarios where isolated, cloud-hosted agents run independently with preloaded credentials, making them suitable for products or features serving multiple clients. Anthropic recommends managed agents for long-running tasks, secure sandboxed cloud infrastructure, eliminating the need for user-maintained infrastructure, and persistent file systems that retain conversation history across sessions.
Managed agents excel in handling workflows that require extended runtimes, unlike skills which are limited to shorter chains of tasks and terminate when the user’s device is off. The cloud-based containerization offers secure, repeatable environments with controlled network access and pre-installed packages, providing a scalable and maintainable solution for teams and businesses. This approach removes the burden of building and maintaining agent loops or sandboxes, as Anthropic manages the infrastructure, allowing users to focus on defining tasks and workflows.
The video walkthrough demonstrates creating a support agent using a provided template, setting up environments, and linking credential vaults. Although the example faced errors due to missing Slack channels and credentials, it showcased the agent’s ability to interact with multiple services like Slack and Notion, manage sessions, and provide debugging tools. The interface includes features for interrupting running agents, viewing logs, filtering errors, and monitoring token usage and session metrics, with expectations for further enhancements in observability.
In terms of cost, managed agents are billed based on session runtime, approximately eight cents per hour, with no charges when agents are inactive. This pricing model encourages efficient session management to optimize expenses. Overall, Claude Managed Agents present a powerful solution for businesses and developers needing isolated, scalable, and persistent AI agents in the cloud, while individual users may still prefer simpler skill-based workflows. The presenter plans a deeper dive in future videos to explore building advanced applications using this new capability.
