Anthropic launches Claude Managed Agents to simplify deployment
Anthropic introduced a new API suite that allows developers to define and start agent sessions without building custom infrastructure. The platform separates execution logic from durable state to improve security and latency.
Anthropic has released Claude Managed Agents (CMA), a suite of APIs designed to lower the barrier to entry for agentic development. The platform allows developers to define agents and start sessions without constructing custom harnesses, effectively abstracting away the underlying infrastructure. CMA shares its sandboxing fleet with Claude Tag but utilizes distinct harness logic to manage operations.
The architecture is built on a separation of concerns, dividing the system into a durable server, referred to as the 'brain,' and an execution sandbox, or 'hands.' This design aims to enhance latency, reliability, and security. To protect sensitive data, credentials are stored in a separate vault and injected via egress proxies, ensuring that agents never access sensitive information directly.
Anthropic also detailed 'Dreaming,' a multi-agent system built on CMA that spins up a secondary managed agent to evaluate and write memories for the primary one. The company is updating CMA to expose primitives that enable users to construct their own multi-agent strategies, such as best-of-N or verifier agents. Anthropic advises developers to treat engineering like research, investing in large suites of evaluations to effectively test these agentic systems. The company positions CMA as a way for developers to focus on application logic rather than infrastructure, emphasizing security-by-default through its architectural separation.