I’ve been running a multi-agent workflow for design + coding. Instead of one LLM, I set up 4 agents: Fast Iterating Developer, Test-Conscious Developer, Senior Engineer, Architect.

It’s not just about code — they generate artifacts like design docs, conflict reviews, and structured PR replies.

Coding agents are great — until you ask them to shut down a server.

My async tool server hit a wall when Claude and AMP failed to handle real-world teardown logic.

Read how I debugged it, what the agents missed, and what I learned:

👉 novibe-vibe-coding.hashnode.dev/when-coding-...

Just shipped a post on how I built a custom MCP tool server for Claude Code + Sourcegraph AMP:

🧰 Tool fetches & parses logs
🧠 Agent diagnoses & collaboratively fixes code
🆕 I just review & push

Full write-up: novibe-vibe-coding.hashnode.dev/vibe-coding-...
GitHub: github.com/mstriebeck/g...