n8n Multi-Agent Intent Routing System: Production Practice of Achieving 85% Cost Reduction via Model Tiering Strategy

The multi-agent intent routing system built on n8n intelligently distributes queries to LLM models of different costs via a model tiering strategy, achieving an 85.5% cost reduction while maintaining a 90.7% routing accuracy, and provides a complete offline evaluation framework. The project is maintained by MatruPrasad09, source code available on GitHub (https://github.com/MatruPrasad09/n8n-multi-agent-intent-router), released on May 24, 2026.

In LLM application production, controlling inference costs while ensuring response quality is an eternal contradiction. This project addresses this issue by proposing an n8n-based multi-agent intent routing solution, which solves the balance problem by intelligently distributing queries to models of different costs.

The system adopts a three-tier routing architecture: Entry Layer (Webhook receives requests) → Routing Layer (Groq Llama3.3 70B for intent classification) → Execution Layer (distributes to three-path agents). Three-path design: PathA (support category, Gemini2.5 Flash, accounting for 41.9%), PathB (technical category, Groq Llama3.3 70B, accounting for 43%), PathC (fallback/unknown, Groq Llama3.3 70B, accounting for 15.1%). The confidence threshold is set to 0.75, which is the optimal balance between utilization and security (at 0.75, 99% of queries enter dedicated agents; increasing to 0.85 drops this to 83%).

Cost Savings: Among 86 queries, system cost was $0.0246 vs full GPT-4o cost of $0.1701, saving 85.5%; for 10,000 daily queries, annual savings exceed $6000. Quality Metrics: Routing accuracy is 90.7% (support category F1=0.92, technical category F1=0.90, unknown category F1=0.90); adversarial query suppression rate is 81.3%; average response latency <1 second. LLM Evaluation: GPT-OSS 120B as the standard, relevance pass rate 81.8%, role consistency 61% (technical agent needs optimization).

JSON Consistency: Three-tier strategy (API-level JSON constraints → regex extraction → default unknown path), prioritizing API-level mandatory syntax. Latency Optimization: Routing layer overhead P50=342ms, user responses are synchronous, logs/evaluation are asynchronous. Choice of n8n: POC phase leverages its advantages in visual debugging and Webhook handling; production environment recommends migrating to Temporal or FastAPI+asyncio to handle high concurrency and idempotency.

V2 Plan: Idempotency (Redis deduplication), timeout budget, circuit breaker mechanism, streaming response, orchestration layer migration, role tuning (technical agent consistency to 85%+), adversarial reinforcement (suppression rate to 95%+). Data Privacy: Production requires PII desensitization (Microsoft Presidio/AWS Comprehend), query hashing (SHA-256), layered log retention (raw logs 24-48h, metrics 90 days).

The n8n multi-agent intent routing system proves that not all queries require the strongest model. Through intent classification and model tiering, significant cost reduction is achieved while maintaining high quality. The project provides a complete evaluation framework and limitation analysis, suitable for teams needing to optimize LLM application costs as a reference.