Sandbox

What it means

The harness runs agent actions inside a container or VM with capped permissions, ephemeral storage, and network allowlists; results stream back to the UI.

Why designers should care

Sandbox UX should state isolation clearly (“Not connected to your repo”), show what data entered or left, and offer promote-to-real-environment paths with extra gates.

Example

A data analysis agent runs Python in a sandbox; charts render in chat with a banner “Sample data only—connect warehouse to run on live metrics.”

Common mistakes

• • Marketing “safe sandbox” while agents share the user’s full filesystem.
• • No escape hatch when users need real credentials after prototyping.
• • Sandbox timeouts with no partial results or resume.

Related terms

• Harness: The harness connects the model to file systems, APIs, browsers, terminals, and review gates, enforcing limits on steps, permissions, and side effects.
• CLI (Command-Line Interface): The terminal accepts structured commands (git, npm, curl) that execute on the user’s machine or remote environment, often invoked by the agent via tool use.
• Tool Use: The model emits structured calls (name + arguments) that your app executes, then feeds results back into the conversation for the next step.
• Guardrails: Guardrails include system prompts, classifiers, allow/deny lists, output validators, rate limits, and human review gates applied before or after generation.
• Capability Disclosure: Visible statements of scope, limits, data access, and known failure modes—not buried in terms of service or a one-time onboarding modal.