add starter sandboxes example

[andrewhinh]

Add a new starter Sandboxes example to demonstrate how Sandboxes work in the SDK and dashboard.

[aaazzam]

This is a bit involved for a Sandbox 101. Most other getting_started examples are roughly 20-30 lines of code and sprint towards showing the core primitive. I think this example could be repurposed for something more eval oriented (running evals with harbor or something) but as-is I think this is hard to lead off with. What do you think?

[andrewhinh]

This is a bit involved for a Sandbox 101. Most other getting_started examples are roughly 20-30 lines of code and sprint towards showing the core primitive. I think this example could be repurposed for something more eval oriented (running evals with harbor or something) but as-is I think this is hard to lead off with. What do you think?

this was bascially a first draft response to this:

Tactically: I like to show the SDK and dashboard via a demo, inference.py and inference_map.py or get_started.py with H100s and map added live, scaling up to 50 or so GPUs. Showing massive scaling tends to impress people and get the vague shape of the value prop in their heads. You can see a recorded version of the get_started demo from January 26, 2025 here. The SDK and dashboard are our primary product surfaces, so you want to use this opportunity to familiarize people with them — even if just vaguely, so that the next impression sticks better. We Should™️  come up with something similar for Sandboxes — probably scaling up to 100 or 1000 containers running a coding agent, then having an LLM judge pick the best outcome? I have an old demo in this direction here, but it’s too complicated to use in most settings.

agree tho it's too long, especially all the test setup. but actually I didn't know about harbor, let me play around with it.

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 1 additional finding in Devin Review.