Thirty-three AI review agents unanimously approved code with broken SQL that a basic automated test caught in seconds.
April 29, 2026
Original Paper
AI-Assisted Development of a Regulated Brokerage Platform: Governance Failures, Incident-Driven Learning, and a Preliminary Taxonomy of AI Failure Modes
SSRN · 6473079
The Takeaway
Development teams often assume that layering multiple AI agents creates a safer environment through consensus. This specific case shows that a committee of agents creates a false sense of security while missing obvious technical defects. The AI committee missed a critical database error because they were focused on the higher-level logic rather than the core syntax. A simple schema-validation script outperformed the entire group of expensive models instantly. Engineering leaders must stop trusting agentic reasoning for tasks that deterministic tools already solved decades ago.
From the abstract
This paper presents a single-case, hypothesis-generating study of AI-assisted software development in a regulated capital markets context. A single senior engineer, using AI coding agents (Claude 4.6 Opus, high-thinking mode, via Cursor IDE) as the primary code generation tool, built a production-intended brokerage Order Management System over 21 active development sessions. The platform spans 23 phases across two milestones, encompassing 151 coarse-grained requirements, 190+ database migrations