Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic

Claude Managed Agents provides a production-grade infrastructure designed to bridge the gap between rapidly evolving AI model capabilities and the rigid, often stale, harnesses used to deploy them. By decoupling the agent’s "brain"—the reasoning loop—from its "hands"—the tool execution environment—this architecture improves reliability, reduces latency by up to 90% for P95 use cases, and enables secure, isolated execution. Key features include persistent session logs for observability and memory, vault-based credential management, and "dreaming" processes that allow agents to self-improve through periodic batch analysis of past interactions. Furthermore, the "outcomes" framework introduces a grader agent to verify task success against defined rubrics, ensuring that agents can autonomously iterate until objectives are met. This approach allows developers to focus on domain-specific logic rather than the complexities of infrastructure, session management, and sandbox maintenance.