Tweet fijado
I built a multi-agent system for Claude Code that builds full-stack web apps end-to-end. One command. Five agents. Zero mock data.
I started with the FullStack-Agent research paper (arxiv.org/abs/2602.03798) and used Google's LangExtract (github.com/google/langext…) to extract 169 structured insights from it. LangExtract is a Gemini-powered Python library that pulls structured data from unstructured text. Perfect for turning a dense research paper into actionable patterns.
Those 169 insights became a working pipeline:
Plan → Backend → Frontend → Test → Verify
You trigger one agent, fullstack-dev, and it dispatches five specialized agents in sequence. The backend agent builds real APIs and seeds the database. The frontend agent wires to those APIs, applies design, then screenshot-loops until 10/10. The tester hits every endpoint, runs Playwright E2E, and verifies DB state after every mutation.
The key insight from the paper: 67% of backend failures are fake endpoints that return hardcoded data instead of querying the database. This pipeline catches that by validating DB state after every action.
The fullstack-dev orchestrator: github.com/sami-abdul/cla…
Full repo with more cool Claude Code stuff inside: github.com/sami-abdul/cla…
English
