Intel / article

What Will Remain Human in Software Architecture? A Focus Group Report

Source: arxiv.org
Author: Uwe van Heesch, Olaf Zimmermann, Christian Kohls
Date: Published Sep 28, 2026

Context

A 22-person focus group of industry and academic participants saw human responsibility shifting toward governing how AI builds software. In the researchers' report, participants broadly agreed that architectural decisions, accountability and the rules constraining agents remain human responsibilities.

They described this work as building the surrounding system of validation, organizational knowledge and company standards. They also tied oversight to uncertainty and the cost of changing a decision. That framing gives teams a concrete way to discuss supervision: decisions that are poorly understood or expensive to reverse deserve closer . These are the participants' views about practice, with a recurring concern that accepting AI decisions without understanding them can erode a team's knowledge.

Key takeaways · AI-distilled

In a 22-person focus group at EuroPLoP 2026, practitioners broadly agreed that architectural decisions, accountability and writing architectural remain human work even as AI agents take on architecture tasks.
Participants named engineering as a new discipline: building the system that governs AI-assisted development, including validation mechanisms, knowledge layers and company-specific standards.
The group proposed criticality, the combination of uncertainty and cost of change, as the single criterion for deciding how much human oversight an AI-assisted decision needs.
A recurring worry was cognitive debt, the gradual loss of human understanding of a system when AI-assisted decisions are accepted without real engagement.

Terms in this piece · Glossary

attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.

Read the source arxiv.org