458K Synthetic RL Environments Yield Scaling Laws for Tool-Use Agents

amaarora · x · 2026-10-29

A new preprint studies how data composition in synthetically generated tool-use environments affects LLM agent training. The authors propose SynthAPI, a pipeline that explicitly controls environment diversity along tools, tasks, and databases, generating 458,752 tasks and training 849 models across four sizes to derive a scaling law that predicts downstream tool-use generalization from environment scale and composition.

Key findings:

As data budgets grow, scaling tool and database diversity beats adding more tasks
Higher data complexity yields larger gains
Agents trained per these findings outperform prior-work data on four benchmarks

A practical reference for anyone doing agent training or eval data engineering.

Original post →

458K Synthetic RL Environments Yield Scaling Laws for Tool-Use Agents

More from coding & agent