Tool-use agents fail in a specific way: they learn the shape of the environments they were trained on and then break on real APIs. The obvious fix is more training tasks. A large-scale study from a Google DeepMind team (Andrew Szot, Samira Abnar, Omar Attia, Abbas Kazerouni and colleagues) suggests that is the wrong lever.
The work introduces SynthAPI, a generation pipeline that explicitly controls the diversity of synthetic tool-use environments along three axes: the tools available, the tasks to solve, and the databases the tools operate on. By systematically varying those controls across 458,752 tasks and 849 trained models in four size classes, the authors derive a scaling law that predicts downstream tool-use generalization from the scale and composition of the synthesized environments.
What the law says
The headline finding is a substitution effect. As the data budget grows, scaling tool and database diversity becomes more effective than simply increasing the number of tasks — a second agent talking to the same APIs learns almost nothing new, no matter how many instructions you vary. And increasing data complexity yields larger gains: environments with more moving pieces transfer better than simple ones.
Following those lessons produces agents that beat models trained on data from prior work in large-scale tool-use agent and synthetic environment generation, and the gains hold across four evaluation benchmarks.
This is the same family of lesson the industry keeps relearning about pretraining data — but here it is measured at the level of API surfaces and database schemas, which is where enterprise agent deployments actually live. If your training corpus exercises a hundred functions across three schemas, no amount of prompt variation on those functions produces an agent that generalizes to a new CRM.
The practical read is uncomfortable for teams that scale agent data by scaling task lists: the binding constraint is probably the number of distinct environments in your synthetic generator, and the pipeline you need to build to widen it, not the volume of prompts you can template.
The paper is on arXiv under ID 2610.37648, submitted 27 October 2026.