Step-level guards for tool-using agents are typically benchmarked against tidy, well-formed JSON tool calls. In practice, however, deployed coding agents act through tools like execute_bash and execute_code, where the payload is a raw string rather than structured JSON. That gap means evaluation results may not reflect how guards behave in real agentic coding workflows.
The paper introduces ToolRosetta, which re-renders tool calls to test whether step-level guards maintain their decisions across different surface forms. The title frames this as "surface-form invariance": a robust guard should behave consistently whether the call is presented as clean JSON or as a string payload for a shell or code execution tool.
The abstract is truncated, so details of the re-rendering methodology and experimental results are not available from the source. What is clear is the motivating claim: evaluations that rely on tidy JSON may miss failures that appear when agents act through string-based execution tools.