Sandbox
Sandbox provides isolated code execution for agents. It is designed for controlled computation, deterministic tool behavior, and safer experimentation.
Execution Variants
Sandbox setups can support multiple execution variants (for example eval-focused or Jupyter-backed modes), depending on your runtime configuration.
What You Can Control
- Sandbox activation and restart.
- Variant selection and reconfiguration.
- Runtime status visibility.
- Execution lifecycle with explicit state transitions.
Typical Workflow
- Launch an agent with sandbox enabled.
- Select the variant needed for the task.
- Validate status and readiness signals.
- Run code/tool operations.
- Reconfigure or restart when switching workloads.
Best Practices
- Keep sandbox enabled only for agents that require code execution.
- Use codemode and guardrails together for stronger policy control.
- Monitor status and failures as first-class runtime signals.
Next Steps
- Learn execution orchestration in Codemode.
- Add policy controls in Guardrails.