Run Claude Managed Agents on Vercel Sandbox

Anthropic's Claude Managed Agents can now run on Vercel Sandbox. Anthropic manages the agent loop, including the model, harness, tools, and session state, while Vercel handles the execution environment, so the agent's tool calls run on existing Vercel infrastructure.

Each session runs inside an isolated Firecracker microVM. The same infrastructure backs 1B+ Vercel deployments, providing the security, availability, and performance needed for workloads that touch internal APIs, internal services, and customer data.

The provided configuration includes the following:

• managed agent loop: Anthropic handles the model, harness, tools, and session state
• Vercel function control plane: creates a sandbox for each session
• Firecracker microVM isolation: per-session isolation with millisecond startup
• credential brokering: brokers credentials at the firewall boundary so secrets never enter the sandbox
• deny-by-default egress: restricts outbound communication via a domain allowlist
• low-latency connectivity: low-latency connections to private networks and cloud workloads

A reference implementation can be found in the step-by-step guide and the Vercel Sandbox documentation.