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.
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)