Test your agent in the workbench

The workbench is where you talk to your agent the way a visitor would — but signed in, with the tools an operator needs. Use it to check that a change to documents, directives, or a routine actually lands before it reaches real conversations. You reach it from an agent’s Chat tab.

The workbench drives the assistant chat API — the same route you would call from your own code to talk to an agent — so what you see here is what a visitor gets.

A grounded answer in the Chat workbench about EU shipping timing, with a Sources line and three suggested follow-up questions, and the conversation id shown in the header.

Chat with your agent

Type a question in the box at the bottom and send it. In the screenshot above, the demo agent is asked how long shipping to Germany takes and answers “3–5 business days after roast day” straight from its Shipping policy document — that cited, specific reply is what a healthy grounded turn looks like. When the agent uses retrieval, the answer carries a Sources line you can expand to see the cited documents. Click a citation to open the source document.

Below an answer, the agent may offer suggested follow-up questions. Click one to send it as the next message — a quick way to walk a visitor’s likely path without typing.

A test conversation is a real conversation with its own id, shown in the header next to a copy button. Copy it when you want to look the conversation up elsewhere, such as pulling its transcript or filing an eval. The workbench keeps one running conversation until you clear it; open the options menu (the ··· button) for Clear chat, which starts a fresh conversation, and for Show citations, which toggles whether the Sources line appears.

Chat and History

The header has a Chat / History toggle. Chat is the live conversation. History lists your past operator test chats for this agent, kept separate from the Activity view so your experiments never mix into real customer conversations.

Inspect a turn

Click any message to slide out Turn diagnostics on the right. It shows the same trace the Activity history drawer shows — how the turn was routed, what retrieval ran, and which stages fired. Select a user turn and the panel resolves forward to the assistant turn it produced. When a turn has a full trace, a Flow button opens a larger view of the turn’s path.

Capture a turn as an eval case

When an answer is worth locking in — or worth catching if it regresses — hover the answer and click the flask icon in its action row. That opens Save this turn as an eval: name the case and confirm. Radioso snapshots the conversation and, when the answer is present, seeds the case with an LLM-judge expectation using that answer as the reference, then opens the eval editor. From there you can run it immediately or add more precise checks.

This is the bridge from “that looked right” to a repeatable test. The Evaluate answer quality guide covers what a case checks and how a run works.

Common failure modes

A weak answer right after uploading a document usually means the document has not finished processing yet, not that retrieval is broken. Check the document’s status first.
No Sources line on an answer: either the agent answered without retrieval, or Show citations is off in the options menu.
The flask icon is missing: it appears on an assistant answer on hover, once the turn has a conversation id and a message to capture.
Your test chats do not appear in Activity by design — look under the History toggle in the workbench instead.