Your own agent
The same eval machinery, pointed at your prompt and your cases.
What this can and cannot do, and where your text goes
- Prompt in, text out. One system prompt, one input per case, one reply. Tool-using agents are not supported here — the agent-loops course has those.
- Any text format. The graders compare the reply as text: exact match, contains, or a judge model against your rubric.
- Through Assay to NVIDIA. Requests go via Assay’s server, because NVIDIA does not accept calls from a browser. Your prompt and cases are sent with each request, used once, and not stored.
- Share publishes. Pressing Share saves the prompt, cases and outputs at a public link for 90 days. There is no delete yet — so do not share what you would not publish.
- Do not paste secrets. Or customer data, or anything from production you would not put in a public playground.
Output must equal the expected value, ignoring case and surrounding space.
Nothing has run yet. Edit the prompt and cases on the left, then press Run.