gcp-agent-tool-trajectory-evaluator

This skill provides the specialized Python logic needed to evaluate how an agent uses its tools. Grounded in evaluation_blog.md, it moves beyond "Did the tool run?" to "Were the tools used correctly and efficiently?"

Usage

Ask Antigravity to:

• "Implement Trajectory Precision and Recall metrics"
• "Set up an Order Match metric for my multi-step agent"
• "Add a custom trajectory scorer to my Vertex AI evaluation"

Metric Definitions

1. Trajectory Precision: Measures what percentage of called tools were actually specified in the reference.
2. Trajectory Recall: Measures what percentage of required tools were successfully called by the agent.
3. In-Order Match: Checks if the required tools were called in the correct sequence (even if other non-essential tools were called in between).

implementation Pattern

Refer to scripts/trajectory_metrics.py. These functions are designed to be serialized and passed to Vertex AI via CustomCodeExecutionSpec.