Tool-using LLM agents need large amounts of training and evaluation data, but gathering real execution traces from enterprise APIs is often blocked by access controls, compliance requirements, and cost. A new arXiv preprint introduces EnterpriseSynth, a method that generates such data from OpenAPI specifications alone, with zero execution.

The paper argues that existing approaches depend on running APIs to collect traces, which is unsafe or infeasible in many enterprise settings. EnterpriseSynth instead synthesizes supervised fine-tuning (SFT) and evaluation data directly from the API specs, potentially allowing teams to build and test tool agents without touching live systems.

The provided abstract does not include experimental results or comparisons, so the effectiveness of the generated data is not detailed in the source. The contribution is the framing: using OpenAPI specs as a zero-execution source for both training and evaluation data.