Harbor + Docker Sandboxes: Run Safe, Reproducible Agent Evals | Docker

This post details the integration of Harbor, an open-source eval framework, with Docker Sandboxes. It enables running AI agent evaluation tasks within isolated microVM environments provided by Docker Sandboxes. Key technical contributions include the ability to define custom network policies for reproducible evals, allowing tasks to download dependencies during setup but restricting internet access during agent execution. It also highlights the scalability of these evals from local development to Docker Cloud Sandboxes, supporting parallel runs without manual provisioning. The post also mentions upcoming support for GPUs and multi-container tasks.