Voice Simulation

Go to app.futureagi.com → Simulate → Agent Definition → Create agent definition.

The creation wizard has three steps.

Step 1: Basic Info

Field	Value
Agent type	Voice
Agent name	voice-support-agent
Select language	English

Step 2: Configuration

Voice agents require provider and contact details:

Field	Value
Voice/Chat Provider	Select Vapi, Retell, or Bland.ai. Others has no API key or Assistant ID fields, so it cannot use observability or outbound calls
Authentication Method	API Key
Provider API Key	Your provider’s API key
Assistant ID	Your assistant/agent ID from the provider
Country Code	Your country code (e.g. +1)
Contact Number	The phone number for inbound/outbound calls
Inbound/Outbound	Inbound (Future AGI places the calls to your agent; Outbound has the agent initiate calls and requires API key + Assistant ID)

Step 3: Behaviour

Field	Value
Prompt / Chains	You are a helpful customer support agent for TechStore. You assist customers with orders, returns, and product questions over the phone. Always be professional, empathetic, and solution-oriented. Keep responses concise: this is a voice call, not a chat. If you cannot resolve an issue, offer to transfer to a human agent.
Knowledge Base	(optional) Select a KB for grounded responses
Commit Message	Initial voice agent prompt

Click Create. You should see the agent definition saved as v1.

Open the agent definition → Create new version → Configuration, and toggle Enable observability (Requires API key) to monitor your voice agent’s calls. The toggle stays disabled until you fill in both Provider API Key and Assistant ID.

Once enabled, Future AGI auto-creates an Observe project named after your agent. After you run simulations (step 5) or your agent receives real calls, voice provider logs are automatically imported into this project. No SDK setup or manual instrumentation is needed.

To view voice traces, go to Tracing (left sidebar under OBSERVE) and select the auto-created project. You should see each voice call logged with metadata such as call duration, status, and transcript.

Go to Simulate → Personas → Create your own persona.

Voice personas have Behavioural Settings (Personality, Communication Style, and voice-only Accent) and Conversation Settings (voice-specific settings including speed, background noise, and sensitivity sliders).

Create these three personas (select type Voice for each).

cooperative-caller

Section	Field	Value
Basic Info	Name	cooperative-caller
Basic Info	Description	A calm, patient customer who explains their issue clearly and follows instructions step by step
Behavioural	Personality	Friendly and cooperative
Behavioural	Communication Style	Direct and concise
Behavioural	Accent	american
Conversation	Conversation Speed	1.0
Conversation	Background Noise	No
Custom Properties	patience_level	high

frustrated-caller

Section	Field	Value
Basic Info	Name	frustrated-caller
Basic Info	Description	An impatient caller who has tried to resolve this twice, speaks in short sentences and may threaten to cancel
Behavioural	Personality	Impatient and direct
Behavioural	Communication Style	Assertive
Behavioural	Accent	american
Conversation	Conversation Speed	1.25
Conversation	Background Noise	Yes
Custom Properties	patience_level	low

confused-caller

Section	Field	Value
Basic Info	Name	confused-caller
Basic Info	Description	A non-technical caller unsure what information to provide, asks for clarification frequently
Behavioural	Personality	Anxious
Behavioural	Communication Style	Questioning
Behavioural	Accent	american
Conversation	Conversation Speed	0.75
Conversation	Background Noise	No
Custom Properties	tech_literacy	low

Click Create on each. You should see all three personas listed as Voice type with their accent and speed shown as tags.

Go to Simulate → Scenarios → Create New Scenario.

Select Workflow builder and fill in:

Field	Value
Scenario Name	broken-device-return
Description	A customer received a laptop with a cracked screen and wants to start a return. They have their order number but don’t know the return process
Choose source	Select voice-support-agent (Agent Definition)
Choose version	v1
No. of scenarios	20

In the Persona section, leave Add by default on to auto-add all active personas, or turn it off and click Add persona to select specific ones.

Click Create. You should see the scenario listed with the 20 generated test cases attached.

Go to Simulate → Run Simulation → Create a Simulation.

The creation wizard has four steps.

Step 1: Add simulation details

Field	Value
Simulation name	return-flow-voice-test
Choose Agent definition	voice-support-agent
Choose version	v1
Description	Testing return flow with 3 caller personas

Step 2: Choose Scenario(s)

Select the broken-device-return scenario.

Step 3: Select Evaluations

Click Add Evaluations and under Groups, select a group of built-in conversation evals for broad coverage (e.g. Conversation Coherence, Conversation Resolution, and Task Completion). Eval groups are workspace-specific, so the exact name and count of evals in your group may differ from this walkthrough.

Step 4: Summary

Review your configuration and click Run Simulation.

Future AGI places calls to your agent in parallel. Each call runs to completion before the result is logged. You should see the executions grid fill in as calls complete.

Once the run completes, the results page shows three tabs:

• Call Details: per-call transcripts, CSAT scores, and evaluation scores
• Analytics: evaluation score distributions across personas
• Optimization Runs: results from prompt optimization runs

Click any transcript to read the full conversation. Look for turns where the frustrated persona escalated, turns where the confused persona stopped understanding, and whether the cooperative persona reached a successful resolution every time.

For example (illustrative: your run’s actual transcripts and scores will differ), the frustrated-caller call on the broken-device-return scenario scored low on Conversation Resolution: the agent confirmed the order number but never told the caller what to do next, and the caller hung up after asking “so what happens now?” twice with no answer.

Fix My Agent: click the Fix My Agent button to open the diagnostic drawer. The platform analyzes the call transcripts and evaluation scores from this run and surfaces two categories of recommendations:

• Fixable Recommendations, organized into two tabs:
  • Agent Level: prompt and behavior improvements you can apply directly (e.g. missing empathy phrases, unclear escalation paths)
  • Branch Level: domain-specific issues grouped by conversation topic or flow (e.g. return policy gaps, billing confusion). Each recommendation highlights which specific calls are affected, so you can trace issues back to exact conversations
• Non-Fixable Recommendations: system-level issues that require infrastructure changes (e.g. missing integrations, data access limitations), plus a human comparison summary showing where a human agent would have handled the situation differently
• Overall Insights: a synthesis of patterns across all calls

For the low-scoring call above, the Agent Level tab surfaced this recommendation: “Add an explicit next-step statement after confirming the order number (e.g. ‘I’ll email you a prepaid return label within the hour’).” Click Apply Fix to create a new agent version with the recommendation merged into the prompt, then rerun the broken-device-return scenario against the new version. On the rerun, the same persona’s Conversation Resolution score moved from failing to passing, since the agent now states the next step before the call ends.

Optimize My Agent: inside the Fix My Agent drawer, click Optimize My Agent to auto-generate improved prompt variants.

1. Enter a Name for the optimization run
2. Choose Optimizer: select from available optimizers (e.g. Bayesian Search, MetaPrompt, ProTeGi, GEPA, PromptWizard, Random Search)
3. Language Model: select the model for optimization
4. Click Start Optimizing your agent

Review results in the Optimization Runs tab. Compare generated prompt variants and their scores to decide which version to promote.