Run untrusted or agent-generated code in a Modal Sandbox

Documented steps

1. Get or create an App handle to own the sandbox: `sb_app = modal.App.lookup("agent-sandboxes", create_if_missing=True)`.
2. Create the sandbox with an explicit lifetime: `sb = modal.Sandbox.create(app=sb_app, image=image, timeout=600, idle_timeout=120)`. The default timeout is 5 minutes; the maximum is 24 hours.
3. Execute commands and read their output: `p = sb.exec("python", "-c", "print('hello')", timeout=3)` then `p.stdout.read()`. Check the process exit code before trusting the output.
4. Attach a Volume for artifacts the sandbox should keep, and a Secret for any credential the sandboxed code legitimately needs — grant the minimum.
5. For sandboxes that must come up before traffic, define a readiness check with `modal.Probe.with_tcp(...)` or `modal.Probe.with_exec(...)`, pass it as readiness_probe, and await `sb.wait_until_ready()`.
6. Name long-lived sandboxes (`name="session-abc"`) and re-attach later with `modal.Sandbox.from_name(...)` or `modal.Sandbox.from_id(...)`; enumerate with `modal.Sandbox.list()` and label with `sb.set_tags({...})`.
7. Always tear down in a finally block: `sb.terminate()` then `sb.detach()`.
8. Reference: https://modal.com/docs/guide/sandboxes