Daily AI CatchupGoogle's ToolGrad Hits 99.8% Success Rate on Real API Chains
Google Research's ToolGrad flips dataset generation for tool-use training: build verified API chains first, then synthesize user questions, instead of inventing queries and hoping agents find working paths. The method achieved 99.8% success on 16,000 real APIs. A Gemma 3 12B model trained on just 500 examples matched Gemini 2.5 Pro on unseen APIs.
Why it matters
💻 Developer · This is how you fix agent reliability. Stop writing synthetic data; start with real successful API chains. You'll need fewer training examples and get models that actually work on production APIs.
📦 Product · Tool-use agents are now trainable at reasonable scale. If you can collect working API chains, you can train a reliable agent. This unblocks 'bring your own tools' features.
🎨 Design · More reliable tool use means you can trust agents with risky operations. UX can shift from 'confirm every action' to 'agents do this by default'—if you're confident the training worked.
📈 Business · Custom tool-trained agents become cheaper to build. Teams with API expertise can train their own models instead of waiting for frontier labs to add every integration. This accelerates adoption.
🤔 Just Curious · This is a fresh angle on training data. It challenges the assumption that you need human-annotated examples. Real API chains are free, perfect ground truth. Why did it take this long to systematize?
Try this: If you're building an agent-based product, experiment with answer-first dataset generation on your tool ecosystem. Start small: pick 10 high-value workflows, generate verified chains, synthesize minimal training queries, then fine-tune an open model. Compare to baseline on held-out tools.
Sources: ToolGrad: Efficient tool-use dataset generation with textual "gradients"