Authors: S Sabouri, H Hasanbeig, A Sage, S Sinha, C Darling, A Tiwari, K Garms, S Chattopadhyay, S Gulwani Status: Under submission, 2026
Summary
Proposes compiling a successful agent run into a reusable neuro-symbolic program (NSP), following a set of guidelines called FlashScript: actions and data transformations become code, instance-specific values become parameters, and judgment-heavy decisions become LLM calls. On 120 WorkBench analytics tasks with Claude Haiku 4.5, the NSPs produced correct results with identical tool calls on every run, using about 70 times fewer input tokens and about a third of the runtime of a skill-guided agent whose runs were inconsistent. On a composite version of the tasks the skill-guided agent’s accuracy fell to 23% while the NSPs stayed correct and consistent, and the NSPs generalized to 108 unseen instances of the same templates.
Links
Related
Microsoft PROSE — Applied Scientist Intern: co-authored with the PROSE team the internship was spent on. Sumit Gulwani: co-author. Katie Garms: co-author. Souti Chattopadhyay: co-author.