Agents
arXiv: TDAD Behavioral Spec Compilation — 92% Success Rate Compiling Tool-Using Agents from Specifications
A separate TDAD paper (2603.08806) treats agent prompts as compiled artifacts: behavioral specifications are converted to executable tests by one coding agent, then a second agent iteratively refines the prompt until all tests pass, with anti-gaming mechanisms including visible/hidden test splits and semantic mutation testing. Across 24 trials on SpecSuite-Core, the methodology achieves 92% v1 compilation success with 97% mean hidden pass rate and 97% mean regression safety under spec evolution. This addresses the fundamental problem that small prompt changes cause silent behavioral regressions — providing a systematic, measurable approach to agent prompt engineering.
Source
↳ Follow the thread