How-to
Test an Agent
The test view is a chat panel backed by the full execution Trace. Every response can be expanded to see LLM calls, Tool invocations, token usage, latency and cost — the same Trace you get in production.
What you'll learn
- How to open and use the test chat
- What each part of the execution Trace tells you
- How Sessions let you isolate test runs
- How to capture feedback and use voice I/O
Open the test view
From the Agent card click Test. The URL is
/agents/{id}/test. The view has a Session list on the left, a chat panel in the middle, and a collapsible Trace under each response.Run a test conversation
- 1
Start a fresh Session
Click + New Session in the left sidebar. Each Session is isolated — memory, context and history do not leak across Sessions. - 2
Send a message
Type into the input and submit. A status badge on the response shows pending, success or error. - 3
Expand the Trace
Click the response to expand. You see every LLM call, every Tool invocation, the reasoning tree, token counts, latency in ms and per-Run cost in dollars. - 4
Give feedback
Use thumbs up or thumbs down with an optional comment. Feedback flows into Eval datasets and surfaces in Analytics. - 5
Try voice (if enabled)
If voice is on, hit the microphone icon to speak. The Agent transcribes via STT, replies as text, and reads it back via TTS.
Reading the Trace
Each Trace is a tree. The root is the user message. Children are LLM calls and Tool invocations in the order they fired. Look for high-latency nodes, unexpected Tool calls, or Guardrail blocks — these are the same signals you will see in Monitor for production Runs.
Frequently asked questions
- Do test Runs count toward billing?
- Test Runs incur the same LLM and integration costs as production Runs. The per-Run cost is shown in the Trace so you can size pilots accurately.
- Can I share a test Session with a teammate?
- Yes. Sessions are scoped to the Workspace, so anyone with Agent access can open the Session list and inspect a Trace.
- Does feedback affect the Agent immediately?
- No. Feedback is recorded and flows into Eval. The Agent itself is not retrained or modified by a thumbs vote.
- Can I replay a Trace against a new Agent version?
- Yes. From Eval you can rerun any captured input against another version of the Agent and diff the outputs side by side.