# Why Humans Can't Keep Up With AI-Written Code - Tariq Shaukat, Sonar

## Executive summary

The shift to agentic software development has led to an abundance of code (up 2,450% in lines of code), but a scarcity of confidence. Tariq Shaukat outlines five critical limitations—including human review scaling issues, agent struggles with large codebases, error compounding, and LLM unreliability—that prevent the efficient scaling of software factories. He proposes the 'Code Confidence Formula' (Intent, Guide Rails, Verification, and Evidence) as the necessary framework for building trustworthy, next-generation software factories. Sonar's solutions, like Sonar Vortex and SonarQube, aim to engineer confidence by making verification and evidence a first-class citizen in the development lifecycle.

## Key takeaways

- The Shift from Volume to Quality: Modern software development principles (Agile, CI, Kanban) draw inspiration from the Toyota Production System, which prioritizes quality and fixing the system rather than just individual symptoms. The goal is to build factories that deliver quality, not just volume.
- The Problem of Scaling Confidence: While code volume is abundant (2,450% increase in lines of code), the increase in releases is only 30%. This gap highlights the need for new tools and processes to scale from code production to reliable shipping.
- The Code Confidence Formula: To build trustworthy agentic software factories, four components must be engineered: 1) Crystal clear Intent, 2) Guide Rails (to keep agents on track), 3) Independent Verification, and 4) Evidence (proof that is consistent, audible, and repeatable).

## Technical details

- Agentic Limitations: Agents struggle with large codebases; accuracy collapses when translating codebases from 10,000 lines to 50,000 lines. Furthermore, errors compound over long tasks: even with 95% accuracy per task, the overall accuracy can drop to 36% over 20 steps.
- LLM Judge Unreliability: LLMs are unreliable judges. Research shows that depending on the model and situation, between 7% and 17% of verdicts change across identical replications. Challenging the LLM can change its mind 7% to 34% of the time.
- Sonar Solutions: Sonar Vortex provides intent guardrails and verifications in the agentic loop, leading to a 36% reduction in token cost and a 92% reduction in issues in the commit. SonarQube focuses on independent verification, showing that users are 81% more likely to believe AI coding has positively impacted code security.

## Practical implications

- Organizations must shift their focus from maximizing code volume to engineering verifiable quality into their software development factories.
- Implementing structured controls (Intent, Guide Rails, Verification, Evidence) is crucial for managing the risks associated with autonomous AI agents.
- Independent verification tools (like SonarQube) are necessary to provide auditable evidence and build confidence in the time-to-market and security of AI-generated code.

## Topics

Software Factory Design, AI Agents, Code Quality, Build Engineering, LLM Governance, Sonar Source, AIE CODE summit

Source: https://www.youtube.com/watch?v=LQW_Wvqhp04