Should AI Engineers Still Read Code in 2026? The Z/L Continuum — Alex Volkov, ThursdAI
Should AI Engineers Still Read Code?
The Changing Landscape of AI Engineering
- A significant shift in AI engineering occurred in December 2025, marking a turning point that altered the field's trajectory. Evidence of this change can be found at wtfhappened2025.com.
- For the first time, AI models began completing tasks that previously took engineers over 16 hours, indicating a dramatic increase in efficiency and productivity.
- Most engineers no longer handcraft code; instead, they supervise AI agents that generate code. This shift is exemplified by Boris from Anthropic, whose entire codebase is now written by AI.
- Approximately 80% of Anthropic's code is now generated by AI, reflecting a broader trend where GitHub anticipates 14 billion commits this year compared to just 1 billion last year.
Perspectives on Code Reading
- At the conference, two contrasting views emerged: one speaker argued for reading every line of code while another claimed "code is free" and emphasized focusing on prompts and guardrails instead.
- Ryan LePopulaire from OpenAI stated that current models are capable enough to write high-quality code without human intervention.
- He suggested that humans should not worry about implementation details but rather focus on ensuring quality through proper prompts and guidelines.
The Debate Continues
- Mario Zechner expressed concerns about relying entirely on agents for coding tasks, warning that they could compound errors without learning from them.
- He stressed the importance of reading critical lines of code to avoid potential issues in production environments.
The Continuum Concept
- Alex Volkov introduced the idea of a continuum between those who read every line of code (like Zechner) and those who trust agents completely (like LePopulaire).
- Volkov has been tracking changes in AI engineering for over three years and noted how these discussions reflect broader anxieties within the community regarding coding practices.
Task-Specific Approaches
- The continuum isn't about individual preferences but rather task requirements; different tasks necessitate different levels of scrutiny when it comes to reviewing code.
- Both speakers agree on inspecting systems rather than every line; understanding what constitutes critical versus non-critical tasks is essential for effective oversight.
Routing Changes Based on Needs
- A routing table was proposed as a method to determine where attention should be focused based on specific needs—critical paths require thorough inspection while less critical areas may allow more leniency.
Future Considerations
- With advancements like Mythos being developed, there’s an ongoing discussion about whether traditional coding practices will remain relevant or if new methodologies will take precedence.
Conclusion: Balancing Oversight with Innovation
- As capabilities evolve rapidly within AI engineering, it's crucial for engineers to maintain their judgment even as some lines may not need direct oversight anymore.