Playground & execution trace

Opening it

Every agent page has a ▶ Test button in the header, available from any tab. It opens a slide-out chat that runs against the agent's live knowledge, tools, and current settings — this is not connected to any real channel, so nothing here reaches a real customer.

What you get

• A normal back-and-forth chat, with memory that persists across turns in the session, so you can test multi-turn conversations realistically.
• Image attachments, if you want to test how the agent handles them.
• New session to reset and start clean.
• An Activity trace panel next to the chat that visualizes exactly what happened behind each reply: trigger matches, memory lookups, knowledge retrieval (which chunks, and their relevance scores), image retrieval, each model turn (tokens used, tools requested), the actual HTTP tool calls made (request and response), MCP tool calls, any grounding retries, citations, and fallbacks.

Why this is the fastest way to debug an agent

If an answer looks wrong, the trace tells you why without guessing: whether the right knowledge was even retrieved, whether a tool call succeeded or failed, or whether the model just phrased something poorly despite having the right information. Start here before changing the system prompt or the model — most issues are a retrieval or tool problem, visible directly in the trace. See Retrieval & grounded answers for tuning based on what you find.

Not the same as Ops

This is a private sandbox, always available to you, separate from the Ops tab — which is where real customer conversations get triaged and taken over by a human. See Conversations & Ops.