Draft Only, Never Send: Hardening AI-Assisted Workflows in a University Research Lab

Joseph Benjamin Ilagan, Gabrielle Anne Uy, Jose Ramon Ilagan

IEEE Software, 2026.

Download PDF (accepted manuscript) Published version (IEEE)

Abstract

AI-assisted workflows often start as open-ended prompting and end up as deterministic scripts. This report describes that shift in two workflows we built in a small university lab in early 2026: a geospatial classification pipeline over about 3,000 establishment records, and a reimbursement workflow that once sent an email we did not mean to send. One hardened because of cost, the other because of risk, but both the same way: we used a capable model to write explicit checks and guardrails, then let cheaper models or plain scripts run inside them. In software-engineering terms, the result is an execution harness around an unreliable component. That construct is old news. What we think is worth reporting is how the harness came together after the fact, in a lab that never meant to build one, and which failures are specific to a multi-model setup, where knowing which model actually acted matters for forensics and is almost never logged.

Keywords: agentic workflows; AI hardening; agent harness; multi-model routing; LLM deployment; software engineering practice

About this copy

Accepted manuscript. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. This is the accepted manuscript of an article accepted for publication in IEEE Software. The final version is available at https://doi.org/10.1109/MS.2026.3716868.

Cite as

Joseph Benjamin Ilagan, Gabrielle Anne Uy, Jose Ramon Ilagan (2026). Draft Only, Never Send: Hardening AI-Assisted Workflows in a University Research Lab. IEEE Software. https://doi.org/10.1109/ms.2026.3716868