AI-generated code can look deceptively clean, well-formatted, well-commented, and test-covered while becoming a serious long-term liability. The danger is large volumes of code that nobody understands well enough to modify safely, explain clearly, or own confidently after it has been merged. This talk discusses the dangers of dark code: code accepted into a production system without genuine comprehension by the engineers who reviewed or approved it.

Dark code becomes more dangerous in agentic development because AI tools produce output that appears more mature than it is. The code can include summaries, comments, generated tests, and convincing explanations, creating the illusion that the team understands what was produced. If the submitting engineer can’t explain what the code does, why it was built that way, what it touches, and what constraints shaped it, the team may have gained speed while quietly creating hidden risk.

This session gives engineering directors, managers, and senior engineers a practical framework for preventing dark code from entering their systems. It focuses on engineer accountability, comprehension as a merge requirement, documented rationale, traceable decisions, and the need for future engineers and agents to infer intent from the code itself. AI can generate the code. Engineers still own what enters the system.

This talk directly references the author's Engineering Standards for Agentic Software Development, especially AC1, AC2, DN2, DN3, and DW2: https://www.linkedin.com/pulse/engineering-standards-agentic-software-development-edensoft-park-ki1se

Target audience:
Engineering managers, engineering directors, senior engineers, staff engineers, architects, and technical program managers responsible for maintainability and ownership.

Preferred Session Duration:
50 mins including Q&A