Test Console

Talk to your agent before any caller does. The voice tray dials the published version while the typed tray runs the draft — the screen says which is which. Beside the conversation, an inspector tray shows exactly what the agent did on every turn.

At a glance

  • The draft by text, the published version by mic
  • A live audio meter and a runtime log under the call
  • Tool calls, evidence and prompt layers, inspectable
  • Per-stage timing, honest when a stage never ran
Test Console

The voice and typed trays

Talk to This Agent starts a call to the published version: while live, the button becomes End with Mute and an audio-level meter, a state line reads where the call stands, and a runtime console underneath keeps the last 200 log lines behind Copy all, Clear and Show or Hide. The typed tray wears a Testing-the-draft-version chip: the thread, then a composer holding Send and — once a conversation exists — Restart. While a turn is in flight the box goes read-only, never disabled, so a follow-up typed mid-send is not deleted.

The Activity tab

One accordion per turn, in order: agent switches first, then every tool call with its arguments and result JSON, then the evidence behind the answer. Each piece of evidence carries a confidence score and an Open source link straight into the knowledge base. A link can open the tray on a chosen tab — an unknown one falls back to Activity.

Prompt and Timing

The Prompt tab shows the prompt layer by layer and the whole assembled prompt in one piece — the left-out layers named too, so you see what the agent was told and what it was not. Timing tells how long each stage took — retrieval, classifying, looking up, searching, the reply and the whole turn — and a stage that never ran reads "not asked", never zero.

  1. Start a call or typeStart a Call dials the published version — the button becomes End with Mute while live. The typed tray runs the draft, and its chip says so.
  2. Open the inspectorThe tray beside the conversation has three tabs — Activity, Prompt and Timing — and Activity lists what each turn actually did.
  3. Restart cleanRestart appears once a conversation exists and drops the whole test conversation — the next run starts from nothing.