Anthropic Launches Platform for AI Agents

Anthropic has launched Claude Managed Agents, a new platform designed for managing complex, long-term AI agent tasks efficiently.

The hosted service on the Claude platform allows for comprehensive management of long-running agents, providing essential interfaces to facilitate their operation.

This innovative solution tackles several persistent issues developers have faced with AI agents, including:

• Loss of context, leading to the need for resets;
• Premature termination of tasks due to model limitations;
• Challenges associated with executing long-duration tasks.

For example, the previous model, Claude Sonnet 4.5, would often end tasks early when approaching its context window limit.

Innovative Approach

Claude Managed Agents streamlines background tasks by handling memory management and error recovery, thus ensuring better reliability and performance.

The architecture separates the agent into three core components, moving away from a singular operational environment:

1. Session: Acts as the agent’s long-term memory, logging all actions to bypass context limitations.
2. Control Logic: Manages agent operations, selecting contexts for the model and directing its interactions.
3. Execution Environment: Facilitates the execution of code, file editing, and external interactions.

This modular structure enhances the resilience of the system, allowing individual components to operate independently without compromising the entire session if one fails.

Anthropic positions Claude Managed Agents as a foundational infrastructure designed to adapt alongside evolving AI models, likening it to an operating system for AI agents.

While the company had previously developed a large language model called Claude Mythos, it was ultimately shelved due to considerable security concerns.