A self-directing agent loop that role-plays the other person, watches where you break, and coaches you — until it decides you're ready.
to a parent who's never talked about anything this personal.
a job you're loyal to, without getting guilt-tripped into staying.
telling your kid you're sick — honest, but still keeping them safe.
with someone you still love, without cruelty or false hope.
You rehearse it in the shower — and it falls apart the second they react in a way you didn't script. Practice partners are expensive, awkward, and never play the part convincingly.
Describe the conversation. An agent becomes the other person — defensive, tearful, guilt-tripping, whatever they'll really be — and after every line you say, it tells you exactly where it landed and how to say it better.
An in-character persona whose difficulty adapts to how you're doing — harder when you cope, gentler when you flounder.
A separate pass scores your clarity, empathy, and heat — and catches the moment you escalate or freeze.
A Readiness score climbs across rounds; the loop self-terminates when you can actually walk in.
Not a chatbot with a loop bolted on. Plan → act → observe → self-correct runs every round, and it's on screen the whole time.
model the other person + a difficulty setpoint from your spec.
you speak; the persona replies in character.
score your message: clarity, empathy, heat, escalate/freeze.
adapt difficulty, coach, update readiness. ↺ repeat.
↺ loops until readiness ≥ 80 across 3+ rounds — a stopping point the system chooses on its own.
| Round score | Persona difficulty | Why |
|---|---|---|
| ≥ 75 | +9 harder | you're handling it → build resilience |
| 40–54 | −4 softer | ease off, keep you in the zone |
| < 40 | −11 gentle | you're floundering → scaffold, don't crush |
Readiness is a smoothed signal, so one great round can't declare victory. All of it is pure and unit-tested — identical on stage as in CI.
A coming-out, a divorce, a diagnosis. So privacy is built in, not bolted on — and the sponsor tools do real work.
The demo engine runs fully on-device. Real calls are PII-redacted; high-risk topics can't leave without explicit consent. Conversations are never stored.
Identity-aware, policy-gated egress at the runtime boundary — mirrored by the app's evaluateEgress().
Claude on Bedrock via a SigV4 signing proxy — a first-class provider, creds never in the browser.
Open, low-cost compute for the private inference proxy — no platform lock-in.
The self-correction is the product, visibly, with a stopping point the loop chooses itself.
Judges have parents, kids, and hard talks of their own. They'll feel this one.
"Throw a wrench" hands the wow moment to the audience — no take-it-on-faith.
Live credential-free demo, 37 green tests, CI + auto-deploy. It actually runs.
Speak your lines; hear the persona.
Your patterns across every rehearsal.
Shareable, self-hostable, on open compute.
Everyone has one. Circles is where you get it right — before it counts.