Claude Managed Agents debuts for long-horizon work

Anthropic is productizing the ugly parts of agent infrastructure, not just the model itself. That matters because the winning platform for agents may be the one that makes long-lived, stateful execution boring and reliable.

• –The big shift is architectural: sessions live outside the harness, credentials stay out of the sandbox, and the runtime can be swapped without breaking the agent contract
• –Built-in support for Bash, file ops, web fetch/search, MCP servers, and persistent event history makes this feel like an agent platform, not a thin API wrapper
• –Anthropic is clearly targeting the “hours-long workflow” use case where most DIY agent stacks become brittle, expensive to maintain, or hard to debug
• –The beta-only posture and research-preview features suggest the platform is early, but the interface-first design is meant to survive model and implementation churn
• –This also sharpens the competitive line versus agent frameworks: Anthropic wants to own the production substrate, while others still leave more of the stack to developers