Skip to main content
Back to HUD Frontier / RSI RL Environments Hackathon Gallery

Project spotlight

Approved

Converge

by Converge

Ship a variant. Watch the data. Keep what wins. Repeat forever.

Converge

About This Project

What is Converge

Converge is an autonomous A/B-testing agent that reads your codebase, learns from past experiments, and proposes ranked conversion hypotheses. It writes and deploys the top hypothesis as a feature-flagged variant without manual work.

Before reaching users, AI personas simulate shopper behavior to catch breakage or degradation. Winners auto-merge to production while Converge monitors conversion and guardrails; learnings compound each round so your site improves measurably every week.

Tech Architecture     •    Multi-agent orchestration: Claude Opus (vision, extended thinking) + specialized planning/coding/persona agents in Python around a single JSON run-state     •    Vision planner: Playwright explores the live site, proposes selector-exact hypotheses with annotated screenshots     •    Coding agents: Git sandbox applies variants as control + variant branches with full diffs     •    Deployment: Vercel CLI deploys each branch to isolated live preview URLs     •    Persona swarm: Claude-generated shopper personas drive headless Playwright browsers; sessions recorded to video, outcomes aggregate into conversion/revenue/lift and guardrail verdicts     •    Auto-promotion: winners merge to main, persistent cross-round history feeds next iteration     •    Dashboard: React + Vite + TypeScript (framer-motion, lucide) streams live pipeline via minimal Python HTTP API

Repository

A repository for converge-agents-2

Python70.5%TypeScript15.4%CSS10.3%HTML3.9%
Last commit 3 months ago