Claude Managed Agents now runs on Vercel Sandbox, combining Anthropic's managed agent loop with Vercel's execution infrastructure. Each session gets its own Firecracker microVM with millisecond startup, the same isolation layer behind 1 billion-plus Vercel deployments. Your private APIs, internal services, and customer data stay inside your own infrastructure.
The architecture is worth understanding in detail. A Vercel function control plane spawns one sandbox per session. Credential brokering happens at the firewall level, meaning secrets never enter the sandbox itself. Egress is deny-by-default with a domain allowlist. That combination of session isolation plus credential handling is the part that makes this production-viable, not just a demo.
Vercel published a step-by-step guide with a reference implementation and full Sandbox docs. Read those for the specifics on wiring tool calls to your existing cloud workloads over low-latency private network connections.
[READ ORIGINAL →]