SynthAPI Finds Tool Diversity Beats Task Volume When Scaling AI Agents

A large synthetic-data study suggests that tool-use agents need broader environments as training budgets grow. In a new paper, Andrew Szot, Samira Abnar, Michael Kirchhof, Alexander Toshev, and collaborators introduce SynthAPI, a pipeline for generating executable tool-use environments. The researchers use it to create 458,752 tasks and train 849 models, finding that tool and database diversity eventually deliver larger gains than additional instructions built around existing environments.

Task counts hide the training mix

Tool-use agents are language models that complete multi-step jobs by calling APIs, reading application state, and updating structured records. Examples include booking travel, changing a customer record, or querying a database before modifying it.

Training these systems requires more than prompt-response pairs. Each environment needs working tools, a stateful backend, realistic starting data, and an executable verifier that can determine whether the agent completed the task correctly. Building thousands of those environments by hand is expensive, so projects such as NVIDIA Nemotron and SynthTools have used language models to generate them.

SynthAPI gives researchers separate controls over three parts of that generated data:

• Tool diversity: the number of applications and APIs represented in training.
• Task diversity: the number of distinct instructions generated for each environment.
• Database diversity: the variety of initial data profiles, which determines the states an agent must inspect before acting.

The pipeline uses language models to implement application APIs as Python tools backed by structured databases. It then generates task instructions, initial database states, and executable success checks. Those controls let the authors vary one source of diversity while holding the others comparatively stable.