Maestro Workflow: Multi-LLM Orchestration with Measured Coordination

Core Philosophy (Paper-Based)

This workflow implements findings from "Towards a Science of Scaling Agent Systems":

1. Tool-Coordination Trade-off

• Paper finding: Tool-heavy tasks suffer from multi-agent coordination overhead
• Our rule: Only Claude Code (orchestrator) runs tools (edit files, run tests)
• Sub-agents (Codex/Gemini) provide TEXT ADVICE ONLY

2. Capability Saturation (~45% threshold)

• Paper finding: When single-agent baseline exceeds ~45%, MAS returns diminish
• Our rule: If you're confident about the solution, SKIP ensemble generation
• Ask yourself: "Am I stuck, or do I just want confirmation?"