Agent Orchestrator Template

Autonomous Orchestration Ecosystem

A self-evolving sub-agent management system for Claude Code. Instead of using pre-defined abstract agents, let the orchestrator dynamically create, integrate, and evolve agents based on actual task requirements.
Original concept by @shintaro_sprech

Concept

The system implements an Infinite Evolution Cycle:

Key Features

- Task-driven agent creation: Agents are born from real task requirements, not abstract definitions
- Dynamic integration: Merges agents when synergy improves outcomes
- Continuous evolution: Strong agents evolve through generations, weak ones fade
- Elite promotion: High-performing agents are promoted to elite status
- Lineage tracking: Integrated agents remember their parents, enabling evolution chains

Evaluation-first Decision Skill

The repository also includes evaluation-first-decision, a shared Agent Skill for evaluating consequential decisions before execution.
It defines success criteria and guardrails before recommending an option, separates facts from assumptions, compares the status quo, and returns one verdict: GO, PILOT, HOLD, or STOP.

- Claude Code: /evaluation-first-decision
- Codex: $evaluation-first-decision
- Detailed guide: docs/evaluation-first-decision.md
- Skill index: SKILLS.md

Use this Skill to produce a Decision Card first. When implementation is approved, pass that card to /task so the orchestrator can execute within the declared constraints and stop conditions.
Global installation scripts for macOS/Linux and Windows are included under scripts/.

Usage: Slash Command

Use the /task slash command to activate the orchestration system.

This command triggers the full orchestration workflow:
1. Reads orchestrator.md
2. Scans pool/ for existing agents
3. Calculates coverage against task…