A new entrant called Sanbox dropped on Hacker News today with a straightforward pitch: give AI agents their own isolated, resumable sandbox environments without the DevOps headache. The platform leverages MicroVM isolation to keep agent workloads contained while supporting reusable templates for common workflows.

How It Works

Sanbox uses the OpenCode SDK as its core harness, which means developers get a standardized way to spin up agent environments that can pause and resume on demand. Each sandbox runs in its own MicroVM—hardware-level isolation rather than container-based—which should appeal to security-conscious teams running untrusted or semi-autonomous code. The resumability factor is key here: you can checkpoint an agent mid-task and pick back up later without losing state.

CLI Compatibility

The Sanbox CLI isn't locked into any single AI toolchain. According to the HN announcement, it works natively with Codex, Claude Code, Cursor, traditional CI systems, or just your terminal. This cross-platform approach mirrors a broader trend in the agent tooling space—nobody wants to bet on a framework that only works with one provider. Templates can be shared across projects and reused, which should cut down on boilerplate when spinning up new agent environments for different tasks.

Use Cases

The platform targets developers building complex multi-step agents that need persistence between sessions, teams running automated code review or refactoring pipelines, and anyone tired of debugging stateful agents that crash midway through long-running tasks. The MicroVM approach also means you can run agents with elevated privileges in a controlled sandbox without worrying about them escaping into your host system—useful for agents that modify files, install packages, or interact with network resources.

Key Takeaways

  • Hardware-level MicroVM isolation (not just containers) provides stronger security boundaries
  • Resumable sandboxes solve the statelessness problem plaguing long-running AI agent tasks
  • OpenCode SDK integration offers a standardized harness approach
  • CLI works with Codex, Claude Code, Cursor, CI pipelines, and vanilla terminals
  • Reusable templates enable team-wide consistency

The Bottom Line

Sanbox is solving a real pain point—AI agents that can checkpoint, resume, and run in true isolation—but the space is getting crowded fast. Whether it gains traction depends heavily on how well the template ecosystem develops and whether the performance overhead of MicroVMs stays acceptable for high-frequency agent invocations. Worth watching, especially if you're already bought into the OpenCode SDK stack.