voice-of-agents
Beta users lie. Simulate them honestly.
- Agentic orchestration
- Data science & evaluation
- NLP & information extraction
An MIT-licensed Python library that simulates user research with Claude, then hands the resulting personas to a live browser-based eval harness that navigates your real product. Published on PyPI.
“Beta users lie. Simulate them honestly, then validate the risky findings with the users you still need to talk to.”
The problem
Pre-product-market-fit user research is too slow and expensive, so founders ship by instinct and learn why users churn only after they have churned.
The approach
Four stages — hypotheses and subjects, typed personas, workflows, journey redesign — and then the part that makes it more than a prompt library: the personas feed a live browser eval harness that drives your actual product. It is not a replacement for real research. It is a forcing function for better questions.
How it works
- research/ — hypotheses and subject framing
- design/ — persona and workflow pipelines, gap analysis
- eval/ — browser driver, the research-to-eval bridge, run diffing
- core/ — persona, pain, capability and backlog models
- cli/ — a `voa` command-line entry point
Stack
Python 3.11+, Click, Playwright, Pydantic, the Anthropic SDK, hatchling. An offline cassette mode so the demo runs without credentials.
Where it stands
- Published on PyPI with CI and trusted publishing
- Roughly 34 test files, a recorded end-to-end run, and a non-suppressible synthetic-data notice emitted every session