SUMM

Claims that AI will drastically reduce the number of software engineers are exaggerated and often made by those distant from real production environments
AI code generation is typically limited by pre-existing architectures and the constraints of large codebases
Much of the generated code offers little innovation because most major decisions about code and infrastructure are already made

AI generates code in the same programming languages as humans, but that code still needs to be maintained and run in complex, error-prone production environments
Production systems have emergent behaviors and require engineers to diagnose and fix unexpected issues
History shows technology adoption (e.g., DevOps, cloud) shifts roles but rarely eliminates them; instead, it raises the level of abstraction

"Vibe coding" means letting AI write code without careful review or understanding of its output, which is inadequate for production-grade software
Writing production software requires meeting high reliability standards (e.g., "four nines" uptime) and handling large-scale data and users
The job of a software engineer includes making thousands of micro-decisions about architecture and implementation, not just generating code

Code is merely an artifact of software engineering, similar to how blueprints are not the job of an architect
Writing less code is often better, as each additional line increases maintenance burden and potential for trade-offs
Major architectural choices (e.g., monolith vs. microservices) involve decisions that AI (LLMs) cannot handle, as they only generate patterns and text, not unique, context-specific solutions

Core task of software engineering is changing software safely without introducing failures or reducing user experience
Approaches for safety include deep codebase familiarity, version control, robust testing, type systems, and controlled deployments
AI can potentially help by providing context, but production responsibility and risk management still falls on engineers

Unlike previous innovations, professional engineers are slow to adopt AI code tools, possibly due to doubts about AI's current capabilities
Recent improvements in AI coding quality (such as the introduction of agents) have accelerated tool evolution
For effective AI co-coding, organizations should document standards, use reproducible environments, maintain easy testing, define clear module boundaries, and create concrete, well-scoped tasks

Treat AI like any team member: provide clear instructions and tools, avoid vague tasks, and set clear standards and guidelines
Recognize that perfect code is not usually generated in the first attempt, whether by human or AI; iteration and testing are essential

Code review remains the most critical skill in engineering, more so as AI and agents write increasing amounts of code
Current code review tools are inefficient, and improvements are needed to handle AI-generated changes effectively
Training and hiring should emphasize code review capability over solving coding puzzles

Remember that AI may mimic human responses but is only generating text, not actually understanding intent; evaluate outputs critically
Accept differences in code style unless functionally necessary; focus on rules and guidelines to standardize where needed
Use context files specifying tech stack and guidelines with AI to get more relevant code outputs
Employ a "create-refine" workflow: have AI draft a plan or document, review and edit, then refine iteratively for best results
Productivity increases when focus shifts from forcing AI to match personal style to ensuring code quality and functionality

Despite rapid changes, software engineering jobs are not disappearing, but evolving
Augment and the speaker are available for further discussion at the event

Vibes won't cut it — Chris Kelly, Augment Code