Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation
Geert Trooskens, Aaron Karlsberg, Anmol Sharma, Lamara De Brouwer, [0.3em] Max Van Puyvelde, Matthew Young, John Thickstun, [0.3em] Gil Alterovitz, Walter A. De Brouwer · 2026
Read Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation by Geert Trooskens, Aaron Karlsberg, Anmol Sharma, Lamara De Brouwer, [0.3em] Max Van Puyvelde, Matthew Young, John Thickstun, [0.3em] Gil Alterovitz, Walter A. De Brouwer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-05150.txt.
Description
Trooskens and eight colleagues study compiled AI, where language models make code during the build. After that, workflows run without model calls, trading freedom for stable results and lower risk. They give a system design, a make-and-check pipeline, and a test plan. On function calls it hits 96 percent, breaking even at 17 deals and cutting token use 57 times at 1,000 deals. Code Factory matches Direct LLM at 80.0 percent on fields and leads item finding at 80.4 percent.
Catalog record
| CREATOR | Geert Trooskens, Aaron Karlsberg, Anmol Sharma, Lamara De Brouwer, [0.3em] Max Van Puyvelde, Matthew Young, John Thickstun, [0.3em] Gil Alterovitz, Walter A. De Brouwer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34535 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.05150