Maestro：模拟高效人类开发团队的多智能体应用开发框架

Maestro is a multi-agent application development framework that embeds software engineering best practices into workflows. It simulates high-performing human development teams via three core roles—PM, Architect, and Coder—to automate production-grade app development. Unlike loose model orchestration tools, it acts as a structured 'application factory' with strict process control.

Current AI coding assistants focus on code generation but overlook software development as a system engineering process requiring planning, review, testing, and collaboration. A single model can't replace a cohesive team. Maestro's core idea: organizing multiple AI agents into human-like team structures yields better production-level results since LLMs learn from human behavior.

Maestro defines three roles mirroring real dev teams:

1. PM: Conducts interactive需求 interviews via Web UI (adjusting for user tech background: non-technical/basic/expert), uses existing code context, outputs detailed specs.
2. Architect: Tech lead—converts需求 to tech specs, splits into user stories, reviews dev plans, enforces engineering principles (DRY, YAGNI, test coverage), merges PRs (no direct coding for objectivity).
3. Coders: Execute tasks from story queue—plan development, write code, run auto tests, submit PRs. Each is an independent goroutine, terminates post-task, state stored in SQLite.

Maestro follows a strict workflow:

1. 需求收集: PM generates specs via interviews (or user provides specs directly).
2. Tech Review: Architect reviews/approves specs (iterates if needed).
3. Task Split: Architect splits specs into user stories and assigns.
4. Dev Implementation: Coders plan, get approval, implement features.
5. Code Review: Architect reviews code/tests, merges PRs.
6. State Persistence: Coders terminate; new coders handle next stories. Auto-retry/reassign if coders stall/fail; issues reported via CLI/Web UI to humans.

• Docker Isolation: All agents run in Docker containers (non-privileged user, read-only root FS, no privilege escalation, resource limits for security).
• Multi-Model Support: Native support for Anthropic/Google/OpenAI models + local models via Ollama; recommended heterogeneous combinations to catch errors.
• Knowledge Graph: Captures architecture patterns, design decisions, coding norms as institutional memory (stored in .maestro/knowledge.dot; coders get relevant 'knowledge packs' on startup).
• Offline Mode: --airplane parameter uses local Gitea/Ollama instead of GitHub/external APIs; --sync to push changes online later.
• Hotfix Mode: Bypasses normal workflow for urgent fixes via dedicated hotfix-001 coder.
• Auto Maintenance: Triggers after certain specs—deletes merged branches, syncs knowledge graph, validates docs, scans TODO/FIXME, suggests test coverage improvements.

适用场景:

• Building production-grade apps from scratch.
• Teams needing strict code review/test coverage.
• Developers wanting to automate routine tasks while maintaining quality.

Limitations:

• Must use Docker (no workaround).
• Standard mode requires GitHub access.
• Primarily for single-user local development.

Maestro represents an evolution in AI-assisted software development—moving from simple code completion/generation to simulating real team collaboration. With clear roles, strict quality gates, and persistent knowledge management, it provides a framework closer to engineering practices. For developers prioritizing code quality over speed, this structured approach bridges the gap from 'toy projects' to 'production systems'.