Test across scenarios
Pair a goal with user data, such as order numbers and account details, for every situation worth rehearsing.
Test across scenarios: Docs (opens in new tab)Simulation
Test your agent against every kind of user
before the first real one shows up.
Set a goal, pick a persona, and let simulated users hold full conversations with your agent, text or voice, scored automatically.
Trusted by teams shipping agents in production
What is Simulation?
Rehearse whole conversations before customers have them. A simulated user, with a goal, a persona and the facts your agent must state, holds a multi-turn conversation with your agent. Evaluators score every run and keep the transcript, so you can see exactly where it went wrong.
A scenario is a user with a reason to be there
Give every simulated user a goal, a temperament, the details a real customer would have to hand, and the facts your agent must get right.
Test across personas
Neutral, friendly, frustrated, confused or a custom persona, so de-escalation and clarity get tested.
Test across personas: Docs (opens in new tab)
Create scenarios using AI
Describe your agent and let Netra draft the scenarios, goals and personas to start from.
Create scenarios using AI: Docs (opens in new tab)Check the agent says what the policy says
List the facts a conversation must convey, and every run is scored on whether the agent got them right.
Check the agent says what the policy says: Docs (opens in new tab)Catch the failures that only show up on turn five
Single-turn tests pass agents that lose the thread. Simulated users hold whole conversations, so memory, flow and resolution get tested the way customers use them.
Run multi-turn evals for text and voice
Whole conversations with text and voice agents, scored on memory, flow and resolution.
Run multi-turn evals for text and voice: Docs (opens in new tab)
Run against your agent endpoint
Wrap your agent in the BaseTask interface and the simulator drives it turn by turn, with session continuity.
Run against your agent endpoint: Docs (opens in new tab)Run in your CI/CD pipeline
Trigger simulation runs from the Python or TypeScript SDK, so conversations are rehearsed on every build.
Run in your CI/CD pipeline: Docs (opens in new tab)Set how long a conversation runs
Cap the turns per scenario: about five for support, six to eight for technical troubleshooting.
Set how long a conversation runs: Docs (opens in new tab)Score every conversation, automatically
Seven predefined evaluators cover conversation quality and agentic behaviour, each scored from 0 to 1. Add your own for rules only your team knows.
Use a set of predefined evaluators
Five for quality and two agentic, all LLM-as-Judge. A score of 0.6 or higher passes.
Use a set of predefined evaluators: Docs (opens in new tab)Bring evaluators of your choice
Add your own evaluators next to the library when a conversation has rules only your team knows.
Bring evaluators of your choice: Docs (opens in new tab)Know whether the goal was reached
Every scenario has a goal the simulated user is pursuing, so a run ends in a verdict rather than a transcript to read.
Know whether the goal was reached: Docs (opens in new tab)
Get scores per turn and per run
See which turn dropped the score, and how the whole conversation did.
Get scores per turn and per run: Docs (opens in new tab)Debug the turn that went wrong
Every run keeps the conversation, the turn-level scores and the traces behind them, so a failed simulation opens on the turn that failed.
See the detailed conversation view and eval results
The full transcript on one tab and every evaluator's result on another.
See the detailed conversation view and eval results: Docs (opens in new tab)Open the trace behind any turn
Follow a low score down to the spans that produced that reply.
Open the trace behind any turn: Docs (opens in new tab)Confirm the scenario as configured
A third tab shows the goal, persona, data and fact checks the run actually used.
Confirm the scenario as configured: Docs (opens in new tab)Review every run at a glance
Scenarios executed, pass and fail rates, and average cost and latency across all runs.
Review every run at a glance: Docs (opens in new tab)Built for text and voice agents
Drive any agent through BaseTask from Python or TypeScript, whatever framework it runs on, including voice pipelines.
- Python SDK
- TypeScript SDK
- LiveKit
- Vapi
- Deepgram
- ElevenLabs
- Cartesia
- LangGraph
- CrewAI
Use the Python and TypeScript SDKs
Wrap your agent once and run simulations from code or your pipeline.
Docs (opens in new tab)Trace and rehearse voice pipelines
Voice agents built on LiveKit or Vapi, with speech from Deepgram, ElevenLabs or Cartesia, land in the same traces.
Docs (opens in new tab)Cookbook
Rehearse the hard conversations first
Cookbook · Simulation
Simulate customer support conversations
Multi-turn scenarios with goals, personas and fact checks, run against a real support agent — catching the incomplete resolutions single-turn tests miss.
Open the cookbook (opens in new tab)What you'll build
- 01Select evaluators for goals, facts, completeness and guidelines
- 02Create a multi-turn scenario with a goal, user data and fact checks
- 03Add scenarios with other personas and intents
- 04Wrap your agent in the BaseTask interface
- 05Trigger the simulation and read transcripts, scores and traces
Frequently Asked Questions
Everything You Need to Know About Simulation
Can't find the answer here? The Simulation docs go deeper, or talk to our team.
What is Simulation?
A way to test agents with realistic multi-turn conversations. You define a goal, a user persona and the facts the agent should communicate, and a simulated user holds the conversation while evaluators score it.
Which personas are available?
Neutral, Friendly, Frustrated, Confused and Custom — from a baseline user to one designed to test de-escalation or clarity, or one you describe yourself.
How are simulations scored?
Seven LLM-as-Judge evaluators — five for quality and two agentic — score each conversation between 0 and 1. A score of 0.6 or higher passes.
How many turns should a scenario have?
Around five turns suits most support conversations; technical troubleshooting usually needs six to eight.
How do I debug a failed simulation?
Open the scenario under Simulation → Test Runs and use its three tabs: Conversation for the transcript, Evaluation Results for the scores, and Scenario Details to confirm it was configured the way you intended.
How does Simulation differ from Evaluation?
Evaluation scores single request-and-response test cases. Simulation scores whole conversations, where the agent has to remember context, handle a persona and reach a goal over several turns.
Start today
Meet your hardest customer in staging
Write a scenario, pick a persona, and read the transcript before anyone real has the conversation. Free to start.