How do you think application architecture has changed with the advent of coding agents?
Short answer
Coding agents push architectures toward being more modular, well-documented, and automation-friendly, because clear boundaries and contracts dramatically improve how effective and safe AI-generated changes can be.
Observed and expected changes
-
Stronger boundaries and layering
- Teams are more motivated to have clean separation between layers (API, domain, persistence, UI) so agents can safely work within a layer without breaking others.
- Clear interfaces (e.g. ports/adapters, repositories, services) make it easier to ask agents to “implement this interface” or “add a new adapter” in a predictable way.
-
Emphasis on contracts and tests
- Public APIs, domain boundaries, and invariants are increasingly enforced with tests, schemas, and types.
- This gives coding agents a reliable contract to follow and makes large-scale, automated changes much safer.
-
More declarative and configuration-driven designs
- Infrastructure, pipelines, and even parts of business logic move into declarative formats (IaC, config, rules) that agents can manipulate in a controlled way.
- This reduces the amount of “hand-written glue” and makes systems easier to reason about for both humans and AI.
-
Architectures optimized for change, not just runtime
- We’ve always cared about runtime performance and reliability; with coding agents, changeability becomes a first-class concern.
- Modular monoliths, well-structured service boundaries, and cohesive domains make it possible for agents to implement new features or refactors without touching unrelated parts.
-
Better documentation as part of the architecture
- Since agents rely heavily on in-repo context, teams invest more in ADRs, architecture diagrams, and inline docs.
- This documentation becomes part of the “surface area” that both humans and AI use to make correct design decisions.
Overall, coding agents don’t replace good architecture; they reward it. The better your boundaries, contracts, and tests, the more your architecture can safely benefit from AI-driven speed.
Interview angle
- “Has AI changed how you structure code?” - toward clearer boundaries and explicit contracts, because both humans and models navigate a codebase better when modules have obvious seams. Typed interfaces and small modules pay off more than they used to.
- “Do you let a model make architectural decisions?” - no. Architecture depends on team, roadmap, existing systems and operational constraints the model can’t see, and the decisions are expensive to reverse. Use it to enumerate options and challenge your reasoning, then decide yourself.
- “What about documentation?” - it matters more. A model reading your codebase benefits from the same context a new joiner does, so a good README and clear module docstrings now have a second consumer.
- “Any downside you’ve noticed?” - it’s easy to generate a large plausible change quickly, which pressures review. Keeping diffs small is a deliberate counter-practice, not a preference.