How to use the Agent Studio – Product Guide

By Alex Bisbe. June 29th 2026

Beta · Admin only. Agent Studio is in Beta and available to admin users. This guide assumes you already know what Agent Studio is — if not, start with the Agent Studio reference page.

This guide is about technique. The reference page tells you what Agent Studio can do; this one shows you how to ask for it so you get accurate, useful answers the first time — and how to chain those answers into real day-to-day work.

The one habit that changes everything: be specific

Agent Studio works against your live data, so vague questions get vague answers. The fastest way to improve a result is to add the three things the assistant can't guess: scope, time range, and the shape of the output you want.

Compare:

Weak: "How are my campaigns doing?"

Strong: "For the last 30 days, show total calls and outcome breakdown per campaign in a table, sorted by call volume."

The second prompt names the period (last 30 days), the scope (per campaign), and the format (a sorted table). You'll get exactly that, instead of a generic summary you then have to refine.

A simple template that works for almost any analytics request:

[What] for [which campaigns/agents] over [time range], as a [table/chart/list], sorted by [metric].

Work in a thread, not in one-shot questions

Agent Studio remembers the current conversation. The best workflows start broad and drill down, each question building on the last:

1. "Show me metrics for the last 30 days across all campaigns."

2. "Now focus on VOICEB.AI - WEB — why is the average duration only 31 seconds?"

3. "List the 10 most recent Dropped calls for that campaign with their transcripts."

4. "Export those to Excel."

Because the assistant keeps context, you don't repeat the campaign name or date range each time. When you start a genuinely new topic, open a fresh conversation (the + in the Conversations panel) so old context doesn't bleed into the new analysis.

Name your conversations. Rename a thread the moment it becomes useful ("Q2 WEB campaign audit") so you can find it again. The Conversations panel is searchable.

Choosing the right model for the job

All three models perform the same actions — the difference is how the answer reads. In practice:

If you want…	Try	Why
A fast, thorough dump of numbers	Gemini 3.5 Flash	Default; detailed and quick.
A tight, skimmable answer for a report	Claude Sonnet 4.6	Concise and well-structured.
A guided "walk me through it" flow	GPT 5.2 Chat	Goal-oriented, step-by-step.

You can switch mid-thread without losing context. A common pattern: explore with the default, then switch to Claude and ask "summarize the key findings from this conversation in five bullets" when you need something to paste into a Slack update.

Five workflows worth saving

1

The Monday performance review

"Compare this week's call volume and conversion to last week, per campaign, as a bar chart. Then list any campaign where conversion dropped more than 10%."

A two-minute health check that surfaces problems before they compound.

2

Diagnose a weak agent before blaming it

VoiceB's rule of thumb: poor outcomes usually trace to traffic quality, not the agent. Use Agent Studio to check that first.

"For [campaign], show the outcome breakdown and average duration. Then list 15 recent Not Interested calls with transcripts so I can see what the incoming traffic looks like."

If the transcripts show off-target callers, the lead source is the issue — not the script.

3

Build a new agent without the forms

"Help me create a new inbound agent for residential energy. Walk me through it one step at a time."

The assistant builds the script blocks (Goal, Script, Objections, Guardrails) with you, one decision per turn. Review each block before confirming.

4

Enrich an agent from a live web page

"Read https://[yoursite]/faq and turn the key points into objection-handling notes I can add to my agent's script."

Great for keeping agent knowledge current with your public pages without copy-pasting.

5

Tidy your call data with tags

"Create a tag called 'Spanish Speaker' that applies when the caller speaks Spanish, and another called 'Pricing Objection' for calls where the customer pushes back on price."

Once tags exist, you can filter and report on them like any other outcome.

Reading charts and tables it gives you

When Agent Studio returns a chart, it also writes key takeaways beneath it. Read both — the takeaways often point you to the next useful question ("VOICEB.AI - WEB is 57% of calls but the shortest duration" is an invitation to drill in). If a chart isn't the format you wanted, just say "show that as a funnel instead" and it re-renders.

Guardrails: this is live data

Agent Studio can make real changes — creating agents, assigning DDIs, editing scripts, setting up tags. Two habits keep you safe:

• Confirm before you commit. The assistant proposes changes; read them before approving. Treat "create" and "assign" prompts with the same care you'd give the dashboard itself.

• Verify numbers that drive decisions. It's an AI assistant and can occasionally be wrong. Before you act on a figure in a board update or a client email, cross-check it in the relevant dashboard section.

Beta means evolving. Capabilities and models will change. If something behaves unexpectedly, start a fresh conversation and rephrase — and check the reference page for the current feature set.

Quick reference: prompt patterns

Analytics    → "[metric] per [campaign/agent] over [range], as a [chart], sorted by [x]."
Drill-down   → "Now focus on [name] and show me [detail]."
Calls        → "List [N] recent [outcome] calls for [campaign] with transcripts."
Build        → "Help me create a [type] agent for [use case], step by step."
Web          → "Read [URL] and turn it into [script notes / knowledge]."
Export       → "Export [that result] to Excel."
Summarize    → "Summarize this conversation's findings in [N] bullets."

Start broad, drill down, confirm before committing, and verify what matters. That's the whole game.