8 agents, 8 repos, one spec — the architecture behind shipping solo with Claude

I'm building a job search tool alone. Backend, frontend, Chrome extension, admin dashboard, e2e tests, marketing site.

The structure: 8 Claude agents, each owning one repo. One spec drives them all.

Every feature starts as a design doc. I brainstorm with the PM and architect agents, then review it myself four or five times. Once the spec is solid, engineering agents work in parallel — frontend and backend don't need to coordinate because the contract is already written.

What actually makes it work:

- Multi-repo boundaries. Each agent only sees its own code.

- Written specs as the single source of truth.

- CLAUDE.md conventions and few-shot examples for consistent code style.

- I review every line of backend code. Frontend I review structure, not pixels.

What I refuse to delegate: product judgment, architecture decisions, and taste. does this code feel like the rest of the system?

Full post with architecture diagrams: https://press.nextloom.ai/2026-05/how-i-build-nextloom-ai-with-claude-agents

What does your AI-assisted workflow look like?