Test agents before they go live.
Script realistic callers, let an LLM play them by chat or real voice call, and let a judge score your success criteria.
Chat or real voice call
Run a fast text chat, or place a real call where a simulated caller talks to your live assistant.
A judge for every criterion
Each success criterion gets pass or fail plus the judge's reasoning, so you see why a run failed.
Response times per turn
Voice runs measure each stage: speech recognition, first token and first audio, as median and p95.
Expected tools checked
Mark the tools the assistant should call. Every run lists its tool calls and checks them.
Tests from real calls
Turn a finished call into a repeatable test in one click. Famulor writes persona, goal and criteria.
Run the whole suite
Run one test or Run all after every prompt or flow change. Milian suggests ready-made personas.
Write the caller once, rerun after every change
Each test stores persona, goal, success criteria, expected tools and max turns. Milian suggests a suite, or you build one from a real call.
- Happy path, edge cases, transfer requests
- Works with prompt and flow assistants
- API, MCP and CLI to create and run tests
Test suite
Persona · goal · criteria
Password reset
ChatLocked-out account holder
3 success criteria
Pass or fail with reasons, not vibes
After every run, a judge LLM scores each criterion and explains why. Results, transcript, response times and tools sit in separate tabs.
- Chat for fast prompt iterations
- Voice for the full speech pipeline
- Expected tools checked on every run
Judge result
Passed · 3 of 3 criteria
- Offers a clear reset path
- Confirms identity first
- Warns not to share passwords
Transcript, reasoning and tool calls stay on the run, so you know why a criterion failed before you ship.
Ship prompts like software. With a test suite.
Open Simulations from the assistant's Test menu. Script the caller, run the suite, read the judge, then publish.
Chat runs are quick for every prompt change. Voice runs place a real call and test the full speech path.
Common questions about Simulations
They can be. In Voice mode, a simulated caller with its own voice from the voice library calls your live, unmodified assistant over the same path a real call takes. Chat mode runs a fast text conversation instead.
Don’t discover prompt bugs on live leads
Build a scenario suite once, re-run after every prompt or flow change, and publish only when the judge stays green.