Google ToolGrad: Auto-Test Your AI Agents (I Tried It)

1 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Oct 3, 2026

Tool Grad from Google Research writes test questions for your AI agents automatically — by running your tools first, getting real results, and then writing questions that fit those results. The answer is already there before the question is even asked.

In this video, I take Tool Grad and use it to build a full exam for a finance agent I created using Google's Agent Development Kit (ADK). The agent can check budgets, pull up expenses, look up vendors, and compare months side by side. Tool Grad generated 10 questions, each using 4 of my agent's tools, and worked out the expected answers too.

I then ran all 10 questions through ADK's evaluation tab and scored my agent — 2 passed, 8 failed. I walk through exactly why most failures happened (tool ordering and incomplete answers) and show how this gives you a clear roadmap for improving your agent.

The Tool Grad paper showed a 99.8% success rate compared to 64% with the old "write questions first" approach. Models trained on Tool Grad examples also outperformed those trained the old way, even beating some big paid frontier models.

If you're building AI agents at work, writing test questions is the boring part everybody skips — but now you can hand it off to Tool Grad.

Chapters
0:00 What is Tool Grad?
0:20 How Tool Grad Works
0:35 Quick Look at the Results
1:14 About Me
1:33 The Tool Grad Paper Explained
2:22 Not a Google Product
2:44 Disclaimer
2:48 Full Demo: Finance Agent Exam
6:04 Why Tool Grad Works So Well
6:15 Researchers vs My Use Case
6:27 Key Takeaway for Agent Builders
6:39 Wrap Up