Architecture Overview · Research & Architecture Explorations

Multi-Agent Orchestration via Workflow Engine

n8n-based 8-agent orchestration system with a central "Walt" orchestrator routing to 7 named specialists across guest experience and SRE domains. Two custom FastAPI microservices provide park tools and BM25+vector log search. Archived at v7.

n8n · 8 Named Agentsv7 Archived · Informed ProductionBM25+pgvector Hybrid RAG

Click any component to explore it in the full case study.

Input A
User Messages (Social Channels)
Messages from X, Instagram, TikTok routed to central "Walt" orchestrator agent. Guest experience queries, park questions, dining requests, and navigation needs.
Social channels → Walt orchestrator
Input B
GCP Cloud Logging (SRE)
BigQuery centralized_logs_dataset. Two-step log retrieval: analyze_* tools return cache_key, then get_cached_logs() fetches full results from Redis cache.
BigQuery logs → Redis cache
8-Agent Orchestration Network — Walt + 7 specialists
Walt
Central Orchestrator
Receives all messages. Routes to sub-agents via n8n sub-workflow invocation based on intent classification.
›
Ops + Weather
Roy + Donald
Roy: downtime/wait times/operational status. Donald: weather forecast and storm alerts.
›
Experience
Tink + Kronk + Wayfinder
Tink: events/activities. Kronk: dining recommendations. Wayfinder: park navigation and routing.
›
Logistics
Jim + Force
Jim: gift delivery orchestration. Force: pre-trip planning and preparation workflows.
Custom AI Agent Service
FastAPI microservice providing 16 park tools. BM25 + pgvector hybrid search for park knowledge retrieval. Redis Stack for job queuing.
FastAPI · pgvector · Redis Queue
n8n-agent-backbone.html
SRE Log Search Service
FastAPI microservice for 12 SRE tools. Runbook search (hybrid BM25+vector), GCP incident status, per-service log analysis. Two-step cache pattern for BigQuery result reuse.
12 SRE tools · BigQuery → Redis
n8n-agent-backbone.html
⚠️ v7 Archived — Docker Compose Complexity
Abandoned at v7 due to Docker Compose complexity becoming unmanageable. Multi-agent persona concept and hybrid BM25+vector runbook search were validated and informed the LangChain production architecture. Named agent pattern (Walt/Roy/Donald) lived on as a design pattern.
['Abandoned at v7 (Docker Compose)', 'Multi-agent pattern → validated', 'BM25+vector hybrid → carried forward', 'Named agents as concept', '8 agents · 2 custom microservices', 'Informed production LangChain design']
n8n-agent-backbone.html#multi-agent cast
8
Named agents
2
Custom microservices
16+
Park tools + 12 SRE tools
v7
Final iteration
Stackn8nDocker ComposePostgreSQL + pgvectorRedis StackFastAPICaddy (SSL)GCP BigQueryGemini AIBM25 hybrid search