Find where your agent breaks — before your users do.
Connect your agent and set it loose on a city of thousands of Census-grounded people. Every conversation is logged, scored, and reproducible — so you find the people your agent lets down, not just the happy path.
Not a few QA scripts. A whole living population.
A handful of QA testers and a small generated dataset only cover the people you already thought of. Synthetic Signals gives your agent a whole city — each citizen with their own identity, personality, and needs, grounded in real data and tailored to your use case.
Thousands of distinct personas — not the 20 you'd hand-write. Synthesized from US Census ACS data, moving on real San Francisco streets — run your agent against the whole population, in parallel, on demand.
Age, sex, job, income — straight from Census data. On top, each citizen carries a modeled personality — the Big Five (OCEAN) traits — and a mood that shifts through the day.
Break the results down by segment. See who your agent works for, and who it quietly leaves behind.
They remember across sessions. Conversations fold into each citizen's memory — test follow-ups and the long game, not one-shot replies. Memory is part of the seeded state, so every run still resets clean.
Grounded in US Census data (ACS), OpenStreetMap & the American Time Use Survey. Synthetic — no real personal data, no real individuals. Score with our built-in rubric, pull transcripts over MCP, or stream every run to your own stack as OpenTelemetry traces.
Your personas. Your audience. Your metrics.
Customize every part of the test — tune each persona's traits, needs and wants; assemble an audience for the exact job you're testing; and report on the outcomes that matter to you.
Tune every trait. OCEAN personality, demographics, and each persona's needs and wants — build exactly the people you're testing for.
Assemble for a specific job. Loan applicants, parking-permit renewals, first-time buyers — test the exact scenario that matters.
Whatever your agent does, there's a custom audience for it.
Its own personas, its own job-to-be-done, its own definition of a good outcome — here's what that looks like across a few real domains.
Polished on the happy path, brittle with the angry caller. Test your support agent on the stressed parent disputing a fee, the retiree who distrusts the chat, the customer fishing for a waiver — then replay the exact failure until it holds.
Try it in the LabBring the agent you already built.
Point any harness at the city over MCP or a plain REST API, and stream every run back as OpenTelemetry — your framework, your language, your evals. No rebuild, no SDK lock-in.
Claude
Gemini
LangChain
CrewAI
Cursor
Vercel
MistralAny agent that speaks MCP or HTTP can run against the city — these are some of the stacks teams build on.
Find where your agent breaks — before your users do.
Connect your agent and set it loose on a city of thousands of Census-grounded people. Every run logged, reproducible, and scored — with our rubric, your own evals, or streamed to your stack over OpenTelemetry. Start in minutes.