Summary
Anthropic’s Claude Managed Agents can now run with customer-controlled execution layers instead of only Anthropic-hosted infrastructure. Cloudflare and Vercel each published concrete sandbox paths that pair Anthropic’s hosted agent loop with isolated runtimes, tighter network controls, and stronger observability.
What changed
Claude Managed Agents added customer-controlled runtime options, with Cloudflare and Vercel both documenting isolated execution environments, private-service connectivity, and governed sandbox operation for managed agent sessions.
Why it matters
This is a concrete step toward hybrid agent deployment: the model loop stays managed, but execution, secrets, and network boundaries move into infrastructure customers can govern. That makes managed agents more credible for enterprise workloads that need stronger runtime control without giving up a hosted agent surface.
Evidence excerpt
Anthropic says Managed Agents can run in customer-controlled sandboxes and connect to private MCP servers, while Cloudflare says its integration gives developers more control over sandboxes, private-service connectivity, and observability.