Evaluate AI Agent Tool Usage Accuracy in Multi-Agent Workflows

This workflow is designed for AI developers building multi-agent systems in n8n. It leverages the built-in Evaluation Trigger and Evaluation nodes to verify if AI agents correctly utilize expected tools. Triggered by chat input or dataset rows, it routes through a multi-tool agent powered by top LLMs, equipped with tools like web search, calculator, vector search, and summarizer. The system extracts intermediate steps (actions + observations), logs actual tools called, and compares them against ground-truth expectations from your dataset. Matches result in a 'pass'; mismatches are flagged as 'fail' with detailed metrics persisted to Google Sheets for debugging.

Key benefits include plug-and-play observability for agent behavior, quantitative performance metrics (e.g., tool_called = true/false), and scalable dataset-driven testing. It saves significant time by automating validation that would otherwise require manual inspection of agent traces, enabling faster iteration on autonomous agents.

Ideal use cases: Testing agent reliability in production-like scenarios, benchmarking different LLMs or tool configurations, and integrating into CI/CD pipelines for AI workflows. Setup involves connecting Google Sheets credentials, configuring LLMs and tools (including Qdrant vector store), and populating your test dataset with expected tool usage per query.