Anthropic PBC today launched Claude Managed Agents, a cloud service that customers can use to build artificial intelligence agents.

The company says the offering shortens the development workflow from months to weeks.

Deploying a production-grade agent requires software teams to build not only the agent itself but also a significant amount of scaffolding. Developers must configure a container in which the agent can perform tasks without risking production systems. Additionally, they must set up the infrastructure on which the container will run, observability features and other components.

Claude Managed Agents automates many of those tasks. It’s accessible through a set of application programming interfaces. According to Anthropic, customers will be billed for the Claude model usage of their agents plus a fee of eight cents per agent runtime hour.

Developers can start a Claude Managed Agents project by providing a description of tasks they wish to automate with an AI agent. From there, they must specify the tools, or third-party applications, that the agent should use to perform the specified tasks. They can also define cybersecurity rules to regulate details such as whether a tool can be activated without user permission.

The agents automatically spinsup an isolated container for each agent. That container hosts software components specified in advance by the agent’s developer. For example, a company building a website design agent might wish to equip its sandbox with a browser.

Claude Managed Agents automates much of the work involved in state management. That’s the task of managing the data used by AI agents to complete tasks. A coding assistant, for example, might incorporate programming advice from the public web into its prompt responses. State management also encompasses more sensitive data such as the login credentials with which an agent signs into cloud-hosted tools.

Tool orchestration is another task that Managed Agents promises to ease. When an agent receives a prompt, the service determines which of the tools at its disposal should be used to generate an answer. An error recovery mechanism enables agents to pick off where they left off after an outage interrupts their work.

Two of Managed Agents’ features are currently in research preview. The first enables an agent to spin up other agents when working on complex tasks. The other feature configures Claude to automatically refine prompt response quality. The company says the latter capability “improved outcome task success by up to 10 points over a standard prompting loop” in internal testing. 

Claude Managed Agents’ initial user base includes high-profile tech firms such as Notion Inc., Rakuten Group Inc. and Asana Inc. Anthropic says several customers have already integrated AI agents built using the service into their products. 

Image: Anthropic