The first time most agents get truly tested is by a real customer. That's a risky way to find out your booking flow breaks on an odd date, or your agent doesn't know how to say no to a refund it isn't allowed to give. Simulator Studio moves that discovery earlier. It builds an AI persona that calls or messages your agent exactly like a customer would, on voice or on text, and puts it through every scenario you care about — automatically, before anyone real is on the other end of the line.
Pick a channel. Pick the agent you want to test. Describe who's calling and how they behave. List the scenarios and what "passing" means for each. Mihu takes it from there — building the simulation agent and running the whole test automatically, then handing you a report of exactly what worked and what didn't.
What Simulator Studio is
Simulator Studio is Mihu's built-in testing ground for AI agents. Instead of picking up the phone yourself, or asking a colleague to try a few messages and hope they remember to check everything, you describe the test once and Mihu runs it for you — with an AI persona standing in for the customer.
That persona isn't a script. It's a simulation agent built specifically for the test: it holds a real conversation with your agent, adapts to what your agent says, and tries the scenario the way an actual caller or texter would — including the awkward, off-script moments that break weaker setups. When it's done, every test case is checked against the success criteria you defined, and you get a clear pass or fail with the reasoning behind it.
Setting up a test, step by step
The setup follows one straight path, and you never have to write a script or touch code.
1. Select a channel
Choose voice or text. On voice, the simulation agent places or receives an actual spoken call with your agent. On text, it carries out the conversation over chat — same rigor, the channel your customers actually use.
2. Select the assistant to test
Pick which of your live agents is under test — your service desk agent, your booking agent, whichever one you're about to ship or just changed.
3. Describe the simulator persona
Who's calling, and how do they behave? A price-sensitive shopper. A confused first-time caller. An impatient customer who interrupts. The persona shapes how the test conversation actually unfolds.
4. Lay out test cases & success criteria
List the scenarios, stacked one after another, and attach what "success" means to each — confirms the right date, declines correctly, stays under a turn limit, switches language cleanly. These are the checks the report will score against.
Creating the simulation agent & running it
Once the four steps are set, Mihu creates a simulation agent built to carry out exactly that persona and those scenarios. From there it runs automatically — no need to trigger each test case by hand, no need to sit and listen to every call. The simulation agent works through the full list, one test case at a time, and every conversation is checked against its success criteria as it finishes.
Set it up once, it runs itself
You describe the test. Mihu builds the simulation agent and runs the whole thing automatically — on voice or text, across every case on your list — and only comes back to you with the results.
Reading the report
When the run finishes, you get a clear picture: an overall pass rate for the test, and the detail behind every individual case.
Every result comes with the reasoning behind it and the relevant part of the conversation, so a failure isn't just a red mark — it's a clear pointer to what to fix.
Let the Assistant design the test for you
You don't have to build the test plan yourself. Ask the Mihu Assistant to test an agent, and it inspects that agent's setup, proposes a relevant persona and a full set of test cases with sensible success criteria, runs the simulation, and comes back with a detailed report — the way a thorough QA colleague would, without you writing a single scenario.
From the Assistant or from MCP
Simulator Studio isn't limited to the dashboard. The same testing capability is reachable through Mihu's MCP server, so external AI assistants like Claude or ChatGPT can trigger a simulation as part of a wider workflow — for instance, running a test automatically the moment a new agent is deployed, without anyone opening Mihu at all.
From the Assistant — inside the app
- Describe the test in one sentence
- The Assistant designs the persona and cases
- Great for a quick check before you go live
- Report lands right back in the same chat
From MCP — part of a workflow
- Triggered by an external AI assistant or pipeline
- Runs automatically on deploy or on a schedule
- Fits into CI-style release checks
- Same simulation engine, different trigger
What teams test
The same simulator shows up differently across industries. A few of the scenarios it's built to catch.
Healthcare · booking edge cases
A simulated patient tries to book on a holiday, asks to cancel twice in one call, and requests a language switch mid-conversation — catching flow breaks before a real patient hits them.
Automotive · service-desk pressure test
An impatient simulated caller interrupts mid-sentence and pushes for a same-day slot that doesn't exist. Success criteria check that the agent stays composed and offers the nearest real alternative.
E-commerce · refund-limit compliance
A simulated shopper asks for a refund outside policy. The success criterion isn't just "declines" — it's declining correctly and offering the right next step every time.
Real estate · lead-qualifying accuracy
A simulated lead gives vague, changing answers about budget and timing. Success criteria check whether the agent still qualifies the lead correctly and captures the right details.
Hospitality · multi-turn reservations
A simulated guest changes party size and dates twice in the same call. Success criteria check the final booking matches the last thing the guest actually said.
Financial services · sensitive-data handling
A simulated caller tries to get account details without verifying identity. Success criteria confirm the agent refuses correctly every single time, with no exceptions.
Find the gap before a customer does
Every one of these is a conversation you'd rather have with an AI persona in a test run than discover live, on a call that matters. Simulator Studio makes that testing automatic instead of optional.
Why it changes the work
- Confidence before launch. Ship a new or updated agent knowing it's already been through the scenarios that matter, not hoping it holds up.
- No manual test calls. Nobody has to sit and role-play a customer, in any language, at any hour, to check that an agent behaves.
- Catches what a script would miss. A real AI persona adapts mid-conversation the way a real customer does — interrupting, changing their mind, switching language — not just reading fixed lines.
- Fits into how you already work. Trigger it by hand from Simulator Studio, describe it to the Assistant, or wire it into a release pipeline through MCP.
Simulator Studio is Day 5 of our launch week. Day 1 introduced the Mihu Assistant, and Day 4 introduced the Builder Assistant, the coding layer that creates custom software. Simulator Studio is how you prove that whatever you built — agent, integration, or automation — actually works, before it meets a real customer.
Frequently asked questions
What is Simulator Studio?
Mihu's built-in testing ground for AI agents. It creates a simulation agent that calls or messages your real agent like a customer would, works through a set of test cases on voice or text, checks each one against success criteria you define, and reports the results.
How do I set up a test?
Open the Simulator module, pick a channel (voice or text), pick the assistant you want to test, describe a simulator persona such as a price-sensitive customer or a confused first-time caller, and list your test cases with the success criteria each one should meet. Mihu creates the simulation agent and runs it automatically.
Can I create a test without opening the Simulator module directly?
Yes. You can ask the Mihu Assistant to test an agent for you in plain language, or trigger a simulation from the MCP server so external AI assistants can kick off tests as part of a wider workflow.
Can the Assistant design the test cases for me?
Yes. Ask the Mihu Assistant to test an agent and it inspects that agent's setup, proposes a relevant persona and a full set of test cases with success criteria, runs the simulation, and delivers a detailed report without you having to write the test plan yourself.
Does it test voice agents as well as chat agents?
Yes. Simulator Studio supports both channels. On voice, the simulation agent places or receives an actual spoken call with your agent. On text, it carries out the conversation over chat. The channel is the first thing you choose when setting up a test.
What does the report look like?
Each test case is scored against its own success criteria and marked as passed or failed, with the reasoning behind the result and the relevant part of the conversation. You get an overall pass rate for the run plus the detail behind every individual case.
Find out if it works, before your customers do.
Open Simulator Studio, describe the persona and the test cases, and let Mihu run the whole thing for you.
Try Simulator Studio Built into every account · €30 free credit to start · Voice & text, from the Assistant or MCP