Every conversation. Under examination.
Human simulation for voice agents. Match experienced test agents to the conversation, then examine what actually happens.
Behavioural testing for voice agents.
Inside the evaluation
A specific caller. A measurable condition.
A caller with a changing request
Voice or text conversation
Confirm the latest preference before creating a booking.
Real runs can generate a suite from your agent description, then execute the selected personas and scenarios.
About this example
Define the caller, the task and what success means.
Authored example · 28 seconds. No audio measurements or evaluation scores are claimed here.
Define the caller, the task and what success means.
Authored example, not a live run. No audio measurements or evaluation scores are claimed here.
AgentNet
A profile.
A history.
A better fit.
A professional network for simulated callers. Each agent brings a distinct persona, voice and behaviour. As it takes part in more tests, its experience builds into its profile.
Match the test to an agent’s skills and relevant experience. Let nForge assign the caller, or choose one yourself.
Meet the test agentsLisa
The caller who changes her mind.
“Actually, wait — could we make that Thursday?”
- Behaviour
- Thinks aloud. Revises answers. Second-guesses a choice.
- Useful for
- Context retention · Request changes · Confirmation
- Experience profile
- Tests participated in, scenarios encountered and the results they produced.
- Selection
- Match relevant experience to the next test.
Illustrative profile structure. No test history or scores are shown.
Who calls.
Where it happens.
What must hold.
Three reusable parts. Combined into a test you can run again.
Agent / Who
Select a caller by persona, behaviour and relevant testing experience.
Scenario / Where
Set the conversation stage, context and required state.
Test / What
Define the goal, objections and conditions that must pass.
From intent to a repeatable test
Describe it.
Define it.
Prove it.
Start with a bot description. Generate a suite, then make the expected behaviour explicit in a structured test definition.
- Agent
- Caller who changes their mind
- Scenario
- Appointment chosen, not yet confirmed
- Goal
- Move the request to Thursday afternoon
- Must hold
- Confirm the latest preference.
Do not act on the superseded request.
A domain-specific language (DSL) makes the test explicit and reusable. This shows its intent, not executable syntax.
Keep the gains.
Catch the regressions.
Assertions
Define what must hold.
Baselines
Compare each run.
CI gates
Block failures. Surface warnings.
Voice or text / WebSocket / Twilio / OpenAI Realtime
View sample report