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

Highlights

• Asynchronous AI agents in the development lifecycle cause a 2,450% increase in lines of code and a 240% increase in commits.

• Only 5% of surveyed developers currently run agentic software factories, but 34% expect to do so within a year.

• Humans agree with AI 93% of the time when the AI is correct, but they also agree four out of five times when the AI is wrong.

• Code translation accuracy by AI agents collapses from 91% down to significantly lower levels when codebase size increases from 10,000 to 50,000 lines of code.

• LLM judges change their verdicts across identical replications between 7% and 17% of the time depending on the model.

• Providing proper intent and guide rails in large complex codebases achieves a 36% reduction in token cost and a 92% reduction in issues in the commit.

Timeline

1. 00:00 - 03:09

   Historical Parallels and the Rise of Agentic Coding

   Industrial manufacturing evolution provides lessons for software production, highlighting how shifts from volume-driven production to zero-rework quality principles transformed industries. Modern software development practices originally adopted these manufacturing tenets, such as agile and continuous integration. The introduction of autonomous AI agents disrupts this balance by making code creation abundant and cheap. Developers rapidly scale commit volumes, forcing enterprises to rethink software factory architecture.

   • Software engineering best practices draw from the Toyota Production System principles of quality over volume and fixing systems.
   • Asynchronous agents generate a 2,450% increase in lines of code and a 240% increase in commits.
   • Survey data indicates that 5% of developers currently run agentic software factories, with 34% planning to adopt them within a year.
2. 03:09 - 10:04

   Core Limitations of AI in Software Development

   Scaling human oversight becomes impossible when code generation outpaces release capacity. Developers exhibit cognitive surrender, deferring to incorrect AI suggestions at high rates. Complex codebases impose heavy token costs as agents spend excessive resources navigating structures rather than writing logic. Furthermore, multi-step tasks suffer from compounding error rates, causing long-running agent workflows to degrade rapidly over successive steps.

   • Human-in-the-loop governance fails to scale because releases only increased by 30% despite massive code volume growth.
   • Cognitive surrender causes humans to accept incorrect AI outputs four out of five times when the AI is wrong.
   • Agent accuracy in code translation collapses when codebases scale from 10,000 to 50,000 lines.
   • Compound errors reduce multi-step task execution accuracy significantly over long running processes.
3. 10:04 - 15:11

   Unreliability of LLMs and the Code Confidence Formula

   Large language models demonstrate volatility as judges, frequently changing verdicts when questioned or re-evaluated. Addressing this unreliability requires a structured formula for code confidence based on four multiplying components: clear intent, active guide rails, independent verification, and repeatable evidence. Implementing these components inside agentic loops optimizes token usage and drastically cuts defects before code reaches commit stages.

   • LLMs acting as judges alter their verdicts in 7% to 17% of identical replications.
   • The code confidence formula multiplies intent, guide rails, independent verification, and evidence together.
   • Sonar Vortex and Sonar Cube implementations reduce token costs by 36% and commit issues by 92%.