I Built a Self-Evolving AI Coding System

Source: Dev.to

TL;DR

I built Vibe‑Claude, a multi‑agent system for Claude Code that:

• Delegates tasks to specialized agents automatically
• Creates new agents when it lacks capabilities
• Learns from failures
• Keeps retrying until done

GitHub:

The Problem

Traditional AI coding follows a loop:

prompt → result → fix → prompt → result → fix …

You still end up doing the manual work.

The Solution

/vibe build a login page with OAuth

Walk away. Return later to a working codebase.

How It Works

Multi‑Agent Architecture

Self‑Evolution

When Claude struggles with a particular task type:

1. A new specialized agent is created.
2. The agent is saved to ~/.claude/agents/.
3. Future tasks of that type are routed to the new agent.

The system continuously improves itself.

Failure Learning

Mistakes are logged to ~/.claude/lessons-learned.md. The same mistake never occurs twice.

Infinite Retry

Attempt 1: Try
Attempt 2: Different approach
Attempt 3: Escalate model
...until success

Quick Start

curl -fsSL https://raw.githubusercontent.com/kks0488/vibe-claude/main/scripts/install.sh | bash

GitHub:

Feedback welcome.