# Harbor Adds Docker Sandboxes to Run Reproducible AI Agent Evals in Isolated MicroVMs

DevFeed: [Harbor Adds Docker Sandboxes to Run Reproducible AI Agent Evals in Isolated MicroVMs](<https://devfeed.tech/articles/harbor-adds-docker-sandboxes-to-run-reproducible-ai-agent-evals-in-isolated-microvms-95268.md>)

Original publisher: [Read original article](<https://www.infoq.com/news/2026/10/ai-agents-harbor-docker/>)

Author: Claudio Masolo

Published: 2026-10-22T09:00:00Z

Content type: news

Language: en

Sources: [InfoQ](<https://devfeed.tech/sources/infoq.md>)

Topics: [Docker Sandboxes](<https://devfeed.tech/topics/docker-sandboxes.md>), [AI Sandbox](<https://devfeed.tech/topics/ai-sandbox.md>), [virtual machines](<https://devfeed.tech/topics/virtual-machines.md>), [Docker Compose](<https://devfeed.tech/topics/docker-compose.md>), [Testcontainers](<https://devfeed.tech/topics/testcontainers.md>)

Tags: [agent-evals](<https://devfeed.tech/tags/agent-evals.md>), [agentic-ai-architecture](<https://devfeed.tech/tags/agentic-ai-architecture.md>), [ai-agents-harbor-docker](<https://devfeed.tech/tags/ai-agents-harbor-docker.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cloud-sandboxes](<https://devfeed.tech/tags/cloud-sandboxes.md>), [coding-agent](<https://devfeed.tech/tags/coding-agent.md>), [compose](<https://devfeed.tech/tags/compose.md>), [daytona](<https://devfeed.tech/tags/daytona.md>), [devops](<https://devfeed.tech/tags/devops.md>), [docker](<https://devfeed.tech/tags/docker.md>), [docker-sandboxes](<https://devfeed.tech/tags/docker-sandboxes.md>), [harbor](<https://devfeed.tech/tags/harbor.md>), [microvm](<https://devfeed.tech/tags/microvm.md>), [news](<https://devfeed.tech/tags/news.md>)

## AI overview

Harbor can run each agent evaluation task in its own Docker Sandbox microVM, isolating it from the host and other tasks. Docker reports testing 64 concurrent cloud tasks; the excerpt gives no data on startup overhead, cost per task, or score variance. Cloud Sandboxes can allow dependency downloads during setup and restrict network access during agent work, while local runs support only public or no-network modes under the machine's existing policy. Docker Compose tasks are supported within one cloud sandbox. The integration is in the nightly release, pinned here to version 0.24.1.dev202610150414, with stable release stated for October 23. Cloud use requires Docker Cloud access and a personal access token with Cloud Sandboxes permission, supplied through SBX_DOCKER_USERNAME and SBX_DOCKER_PAT; select it with `-e sbx`. Local use requires `sbx login` and `--ek target=local`. Network restriction reduces one avenue for reward hacking but does not prevent tampering with writable tests or score files.

## Source excerpt

Docker announced that Harbor, the open-source agent evaluation framework, now works with Docker Sandboxes. This change allows each evaluation task to run in its own microVM. So, instead of sharing a container, tasks are isolated for better performance By Claudio Masolo