Nineteen AI agents, seven stage gates, one terminal tab

What it does AppGenesisForge is a Claude Code Agent Teams scaffold that installs 19 specialized AI roles—developers, reviewers, QA, and a product-lead—into your project. It drives them through seven stage gates (change folder → TDD implementation → code review → UAT → E2E → sign-off) where mandatory skills, blocking hooks, and Definition-of-Done checklists prevent the process from skipping steps. It targets Web (React/FastAPI), WeChat Mini Programs, and Apple native stacks (Swift/SwiftUI), and can switch between DeepSeek, Doubao, Qwen, and MiniMax backends.

The interesting bit The project doesn’t bet on smarter models; it bets on dumber, more reliable process. The “anti-AI aesthetic” design discipline is a nice touch—there’s literally an ADR banning purple gradients, emoji icons, and fake data to stop the UI from looking like a ChatGPT demo. The four-layer hook defense blocks dangerous commands and scans for leaked keys before any commit lands.

Key highlights

• 19 AI roles with enforced separation of duties: reviewers can’t edit their own reviews, QA can’t sign off on their own tests.
• Seven stage gates with a fast lane for small patches, though high-risk changes must pass every gate including SIT-audited review and user-approved UAT cases.
• Real-time HTML dashboard generated locally with zero server dependencies; task cards update roughly every three seconds.
• Multi-LLM wiring for DeepSeek, Doubao, Qwen, and MiniMax, plus dedicated tracks for WeChat Mini Programs and Apple native with notarization workflows.
• “Anti-AI taste” rules baked into the design discipline layer to prevent generic AI-generated UI patterns.

Caveats

• Explicitly excludes Android native, Windows/Linux desktop, and LLM training workloads.
• Even the “fast lane” for small patches still requires deployment and smoke tests.
• The README flags one mandatory human checkpoint: verifying ADR-000 so the AI doesn’t hallucinate your tech stack.

Verdict Worth a look if you’re already using Claude Code and find that single-agent sessions collapse on features longer than a few hours. Skip it if you just want quick AI-assisted snippets or if you build Android/desktop apps.