I built an agent orchestration layer for Claude Code that creates new agents when it encounters capability gaps.
Core features:
- 11 specialized agents (analysis, planning, implementation, UI, etc.)
- Auto-routing based on task type
- Self-evolution: creates new agents/skills when needed, saves to ~/.claude/agents/
- Failure learning: logs mistakes to prevent repeating them
- Infinite retry with model escalation (Haiku → Sonnet → Opus)
The interesting part is the evolution. When the system struggles with something repeatedly, it creates a new specialized agent for that task type. All evolution is logged and the agents persist across sessions.
Uses Opus heavily, so not cheap, but the goal was effectiveness over efficiency.
kyoungsookim•15h ago
Core features: - 11 specialized agents (analysis, planning, implementation, UI, etc.) - Auto-routing based on task type - Self-evolution: creates new agents/skills when needed, saves to ~/.claude/agents/ - Failure learning: logs mistakes to prevent repeating them - Infinite retry with model escalation (Haiku → Sonnet → Opus)