AI Can Write More of the Code. We Still Have to Engineer the System

When discussing AI's impact on software development, it is crucial to clarify which specific aspect the statement refers to. Automating code production is just one facet, while design, decision-making, and testing are distinct activities. Automation in implementation does not negate the need for engineers to determine what should be built and what evidence suffices to accept a system. The responsibility does not diminish; instead, it transforms into a collaborative effort between AI agents and human developers.

One concern is that portraying AI as replacing developers reduces their role to mere coding, overlooking the valuable assistance agents can provide in analysis, design, and testing. This reductionist view undermines the complexity of AI's capabilities and misrepresents the nature of job automation. Evaluating workforce replacement requires more than impressive demonstrations; it demands substantial evidence.

The practical application of AI in development centers on maintaining transparency in decision-making and ensuring useful feedback. Conversations are helpful for exploring options but are fragile as the sole record of architectural decisions. Models have limited context windows, which can lead to loss of context or unclear instructions when sessions are summarized or resumed.

To mitigate these issues, it is recommended to assign durable decisions to a concise, reviewable location, such as a PROJECT_GUIDE.md file. This file is not inherently special but can be connected through the tool's project instructions or explicitly included in the task. Each increment of work should provide relevant files for the agent to interpret and a brief account of the intended change, constraints, and acceptance criteria.

Selective retrieval of documents during tasks can balance context and cost but requires careful consideration. While it may reduce input and potentially increase efficiency, maintaining, retrieving, and correcting documents also consume resources. Any claimed savings need measurement against the entire task. For instance, consider a hypothetical mobile search screen: the user wants to search saved items for easy access without scrolling through the entire collection.

In this example, explicit acceptance cases are crucial. They help clarify technical questions, like whether filtering requires asynchronous work. The agent can assist in evaluating these decisions, but the choice of architecture—such as unidirectional data flow, explicit UI states, and separation between filtering logic and rendering—should be an engineering decision with a clear purpose, documented in the project guide.