Multi-Agent AI Orchestration

Working across multiple projects with AI agents meant constantly re-explaining context, losing knowledge between sessions, and duplicating effort. Each project was an island—patterns discovered in one couldn’t be reused in another without manual transfer.

Tracked repeated pain points across sessions: context loss, knowledge duplication, manual coordination overhead. Researched RAG architectures and multi-agent patterns. Identified that the core problem was knowledge persistence and distribution, not agent capability.

Built a hub-and-spoke architecture: central knowledge hub stores patterns, tools, and decisions in structured JSON. Spoke projects sync from the hub and contribute back. Production RAG v2 pipeline with hybrid retrieval (vector + BM25 keyword search), cross-encoder reranking (BAAI/bge-reranker-base), semantic code chunking via tree-sitter, and 5 configurable index categories. FastAPI endpoints for cross-project search. Hook enforcement system (15 hooks — 3 blocking, 12 advisory) ensures agents follow startup protocols, commit cadence, RAG-first lookups, and session documentation — with file-system-verifiable markers instead of advisory-only reminders. Native Claude Code agent teams with 6 typed roles (orchestrator, implementer, qa-agent, security-agent, researcher, reviewer). Built-in skills (/qa, /security-audit, /self-eval, /decision-review) for structured validation workflows. OODA decision framework (Observe-Orient-Decide-Act) used for all significant changes.

• Hub-and-spoke over monolith — each project stays autonomous while sharing knowledge through a central sync mechanism
• Hybrid retrieval (vector + BM25) — pure vector search missed exact matches like task IDs and port numbers; BM25 fusion catches them
• Cross-encoder reranking — reorders combined results for semantic quality, with configurable top-N and similarity cutoff guardrails
• RAG-first lookup policy — cut token usage by 90%+ by querying indexed docs instead of reading full files
• Native agent teams over manual spawning — typed roles with scoped permissions reduce coordination overhead and enforce separation of concerns
• OODA framework over ad-hoc decisions — mandatory Observe-Orient-Decide-Act loop ensures agents gather context before acting

• Production RAG v2 pipeline with hybrid retrieval, cross-encoder reranking, and semantic code chunking
• Cross-project search enables querying any spoke's indexes from any other spoke
• 15-hook enforcement system (3 blocking, 12 advisory) ensures protocol compliance with file-system-verifiable markers
• RAG-first lookups cut token usage by 90%+ compared to reading full files
• 6 typed agent roles via native Claude Code teams: orchestrator, implementer, qa-agent, security-agent, researcher, reviewer
• Built-in Claude skills (/qa, /security-audit, /self-eval, /decision-review) for structured validation workflows
• OODA decision framework ensures all significant changes follow Observe-Orient-Decide-Act loop
• Automated hub knowledge versioning with drift detection and auto-sync across all spokes