Test an agent in the Playground

The Playground is a chat in the dashboard for trying an agent before you ship it. It uses the same pipeline as the API: knowledge search, tools, streaming and tracing. Each reply costs 1 credit, like any other.

Before you start

• An agent with a working provider key.
• A few credits on your balance.

Steps

1. 1

   Open the Playground

   Choose Playground in the sidebar and pick the agent.

2. 2

   Ask a question

   Type a message (up to 32,000 characters). The answer streams in. Under each reply you'll see its latency, tokens, the tools it called and the sources it used.

3. 3

   Open the details panel

   Beside the chat (below it on a phone) the panel has three tabs. Prompt holds the draft system prompt your next message uses. Tools lists the agent's MCP tools and their parameters, with required ones marked. Log shows what the Playground sent and received, newest last; you can clear it.

4. 4

   Try a prompt change

   Edit the prompt on the Prompt tab. While the draft differs from the live prompt, your messages use the draft. Reset it to go back to the published prompt.

5. 5

   Publish the draft

   When it's right, choose Publish prompt. It becomes the agent's prompt on every channel and is recorded as a version.

The same from code

The API's prompt field replaces the system prompt for one request, so you can test a draft from a script or an eval harness:

{  "message": "Summarise our refund policy in one line.",  "prompt": "You are a terse support agent. Answer in one sentence."}

If something goes wrong