Software Architecture in the Age of AI: Emerging Patterns and Challenges

The Evolving Landscape of Software Development with AI

The integration of AI, particularly Large Language Models (LLMs) and 'agentic programming', is rapidly transforming software development. Early discussions about *what* agentic development might look like have shifted to *how* it's being implemented in production. This shift brings new terminology like "harness engineering" and new concerns, such as the escalating cost of LLM tokens, which were not prominent just a few months prior.

The Enduring Importance of Architecture and Design

A central question is whether traditional software architecture and design remain important when AI agents can generate code. Two main hypotheses emerge: one suggests AI's "Galaxy Brain" can handle any complexity, negating the need for careful design; the other, termed the "Venn Diagram of Developer Experience and Agent Experience is a circle," posits that AI benefits from good design just as humans do. Modularity, clear naming, and good internal design quality (ease of change) are crucial for both human and agent comprehension, potentially even reducing AI token costs for the same change.

• Desirable Outcomes: Evaluate architecture based on outcomes like ease of change over time, not just short-term metrics.
• Token Costs as a Metric: Lower token costs for changes can indicate better architecture, as AI agents process well-structured code more efficiently.
• Mechanical Sympathy for LLMs: Understanding how LLMs work is essential for effective integration, much like understanding the underlying hardware in traditional development.
• AI Amplifies Existing Issues: LLMs trained on existing codebases will amplify their inherent problems if the source code is poorly designed.

New Challenges and Workflows in an Agentic Age

As AI agents become more involved, new challenges and workflows arise. These include managing the quality of AI-generated code, which often exhibits duplication and mixed concerns (e.g., domain and display logic). The article proposes a workflow where agents assist in breaking down stories, creating task lists, and even generating documentation, with human oversight. Another critical concern is the risk of dependence on AI, highlighting the need for good design as a hedge against potential unavailability or prohibitive costs of AI tools.