AI Agent Tool Usage Accuracy Assessment - n8n Workflow

AI Developers and Engineers who need automated testing for multi-agent LLM systems.
Data Scientists focused on measuring and improving model reliability.
Users of n8n seeking advanced examples of evaluation capabilities.
QA specialists ensuring accuracy in complex n8n workflow configurations.

Developing reliable AI agents often requires rigorous testing, especially when the agent has access to multiple functions or tools. This sophisticated n8n workflow addresses the challenge of verifying whether an agent correctly chooses and executes the necessary tools for a given query.

By leveraging the specialized n8n Evaluation nodes and connecting them to an external dataset (Google Sheets), this n8n workflow runs predefined questions against a multi-tool agent (which uses Search Agent, Calculator, Web search, and Search_db). The key value of this n8n workflow is the automated assessment: it compares the list of tools the agent should have called (from the dataset) against the list of tools it actually called (from the agent's intermediate steps).

This makes this specific n8n workflow ideal for continuous integration and quality assurance (QA) in your AI development pipeline. If you are looking for robust n8n templates for agent evaluation, this example provides a powerful foundation.

This process starts with an n8n trigger designed specifically for testing or an interactive chat trigger for live usage.

1. Evaluation Trigger: The primary flow starts with the When fetching a dataset row n8n trigger, which pulls a test case (input question and expected tools) from a Google Sheet dataset called 'Tool calling'.


2. Input Preparation: The data is processed by the Match chat format n8n node to prepare the prompt for the agent.


3. Agent Execution: The Search Agent n8n node executes the query. This powerful agent has access to multiple tools: Calculator, a Summarizer workflow, a Web search tool (using Firecrawl), and a Qdrant-backed Searchdb retrieval tool. The agent uses the OpenRouter Chat Model as its brain.


4. Flow Control Check: The Evaluating? n8n node checks if the n8n workflow is running in evaluation mode (i.e., triggered by the dataset trigger).


5. Tool Call Verification (Evaluation Mode): If the workflow is evaluating, the Check if tool called n8n node executes a complex expression. This expression compares the expected tools (from the dataset row) against the actual intermediate steps recorded by the Search Agent n8n node, resulting in a boolean toolcalled metric.


6. Dataset Output: The Set Outputs n8n node records the actual tools called back into the Google Sheet dataset for later review.


7. Metric Finalization: The final Evaluation n8n node converts the boolean result into a quantifiable metric (1 for success, 0 for failure) and submits the evaluation results, allowing users to track the tool usage accuracy of the agent across many test cases. This powerful n8n node completes the automated test cycle.

To deploy this evaluation n8n workflow, follow these steps:

1. Import: Import this n8n workflow JSON into your n8n instance using the 'New' -> 'Import from JSON' option.


2. Credentials Setup: You must configure the following credentials for this n8n template to function:

1. Link Tools: Ensure the Search Agent n8n node correctly references the four tool nodes (Calculator, Summarizer, Web search, Searchdb) and the language model (OpenRouter Chat Model).


2. Dataset Configuration: Configure the When fetching a dataset row n8n trigger to point to your specific evaluation Google Sheet document and sheet name, ensuring the expected tool names are present in the dataset rows.

When fetching a dataset row (Evaluation Trigger): This n8n trigger initiates the evaluation cycle by fetching a single row (test case) from a Google Sheets dataset. Key configuration includes the Document ID and Sheet Name ('Tool calling').
Search Agent (LangChain Agent n8n node): The core intelligence. It processes the input query, decides which of the four connected tools to use based on the extensive system message, and crucially, is configured to return intermediateSteps which contain the record of tools actually called.
OpenRouter Chat Model (LangChain LM Chat n8n node): Serves as the Large Language Model engine for the Search Agent. This powerful n8n node dictates the agent's reasoning.
Searchdb, Calculator, Summarizer, Web search (Tool n8n nodes): These are the functional tools provided to the agent. The Searchdb uses a Qdrant vector store and requires the Embeddings OpenAI n8n node.
Check if tool called (Set n8n node): This crucial n8n node implements the evaluation logic. It compares $json.toolstocall (from the dataset) against the agent's actual actions in $json.intermediateSteps to generate a boolean success metric.
Evaluation (n8n node): Used twice. First, the Set Outputs instance writes the agent's intermediate results back to the sheet. The final Evaluation instance sets the metric tool_called (0 or 1), completing the automated assessment for this specific n8n workflow iteration.