Upcar is a P2P car-rental marketplace — a Turo competitor, live in 5 states, ~$100K monthly GMV, ~300 hosts. The agent fleet on this site is not theory. Every agent shipped because Upcar needed it that week. The architecture generalizes — the wedge was specific.
Solo founder. ~300 hosts. Five states. Chargebacks at 11pm. KYC escalations from three time zones. Stripe drift on Friday afternoons. Each agent was a thing the founder built because he couldn't be in two places at once.
Then the founder noticed the agents were a moat — not the marketplace, not the brand, not the hosts. Eighteen months of operating context informed how each agent was designed: the shapes, the envelopes, the decision classes, the failure modes the runtime had to handle. Praxis is what happens when you turn that context into a product.
| Q4 2024 | Upcar launches in CA. P2P car rental, ~30 hosts, ~$8K/mo GMV. |
| Q2 2025 | First chargeback wave. Founder writes the dispute-response runbook by hand. Wins 4 of 7. |
| Q3 2025 | Expand to NV, TX. Second KYC rubric needed. Founder writes a per-state policy doc. |
| Q4 2025 | First booking-risk script: Slack bot joins Stripe Radar + Persona + account age. Stops a $4,200 Audi RS5 from a 3-day account. |
| Q1 2026 | AZ + FL. Five state rubrics. Founder realizes the ops playbook is the moat — not the marketplace. |
| Apr 2026 | Anthropic ships SKILL.md. Tom Blomfield writes the YC Company-Brain RFS. The shape of the moat snaps into focus. |
| May 2026 | Praxis spins out. The agents become the starter fleet. Onboarding the first operators. |
Sanitized but real. Trip numbers and dollar amounts come from the actual Upcar history.
payouts-recon caught an asymmetric-drift bug in tripPriceHelpers.ts — a host was overcharged $493 because the rounding step ran on the wrong side of the split. Drafted the fix as a PR. Founder reviewed, shipped to development, refunded the host. Bug caught and refunded on the same day it happened.
dispute-response shipped Stripe CE 3.0 packages on three trips between 11pm and 2am — pulling GPS, check-in photos, host messages. Founder reviewed in the morning. Two won, one lost.
trust-safety flagged 6 accounts targeting Audi/Porsche listings, all from one device fingerprint. Proposed a Sift policy diff. Shipped same week. Zero successful bookings from that cluster since.
pr-reviewer (Claude + Codex tiebreak) caught a security regression in OAuth-callback handling that human review missed. The two LLMs disagreed; the disagreement was the signal. Founder rolled back, fixed, re-shipped.
Embedded clip of the actual Trip 10583 catch — payouts-recon flags the asymmetric-drift bug, drafts the PR, founder reviews + ships.
Your company will produce the same kinds of decisions. Praxis turns them into agents.
Onboard your company →