AI Engineering Is Easy. Changing How We Work Is Hard

Artificial intelligence is becoming an increasingly important tool in software development. Terms like agentic development and AI harness engineering illustrate the growing integration of AI in our industry. However, while AI can assist with requirements, challenge product requirements documents (PRDs), explore UX ideas, reason about architecture, create implementation plans, write code, and validate results, there are still significant hurdles to overcome before it can fully revolutionize how we build software.

The traditional software development workflow—comprising requirements, refinement, planning, build, and validation—can be enhanced by AI in each stage. AI can help clarify problems, question assumptions, identify gaps in PRDs, turn well-defined requirements into plans, and generate implementation tasks. However, AI relies heavily on explicit problem, scope, constraints, edge cases, and expected outcomes provided by humans.

It acts as a reviewer, asking questions about potential failures, testing, contradictions, and overlooked aspects. AI also helps compare different versions of PRDs and enables one model to review another's output, making it easier to spot gaps.

Interestingly, coding may not be the bottleneck anymore. Complex work often involves multiple rounds between product, UX, requirements, and engineering before development can begin. With AI's ability to produce working implementations quickly, waiting for clarified requirements can become a bigger issue. This suggests that engineers should be involved earlier in the process, and UX needs to be part of the conversation from the start.

Lightweight prototypes or wireframes created with AI can expose gaps in requirements much faster than further discussions.

While it may seem intuitive to give AI more documentation to work with, focusing on clear project structure, focused documentation, useful agent instructions, architecture decisions, and a well-organized codebase can be more beneficial. The goal is to make it easy for AI to find what it needs when it needs it, rather than overwhelming it with a plethora of information.

Moreover, the ticket-based approach may not be ideal for AI, as it can be difficult for agents to reason about weeks-long or month-long tasks. Breaking work into manageable scopes with clear definitions of done would be more effective.

Applying the AI workflow to every aspect of software development may lead to a new form of bureaucracy. Not every feature requires the full AI lifecycle; simpler changes might only need a prompt and a developer. The workflow should match the complexity of the work. For AI to be effective, it needs to understand the real system, which means having access to the relevant codebase. This introduces security, licensing, privacy, and organizational considerations, so adoption isn't merely a decision about the right model.

As multiple developers and agents work in parallel, the potential for conflict increases. Each agent has its own context and can make decisions based on what it currently sees. One agent might change something another doesn't know about, or two agents could make decisions that don't work well together. Therefore, good engineering practices, such as small changes, clear boundaries, thorough testing, frequent reviews, and consistent project rules, become even more crucial.

In conclusion, AI isn't ready to completely transform software development, but it can significantly improve our workflows by addressing ambiguity, improving efficiency, and uncovering hidden bottlenecks. We should experiment with these workflows, identify areas where they fall short, and continuously refine the process. AI is fundamentally changing how we approach software development, exposing many aspects that slow us down, and it's up to us to adapt and leverage its potential.