Architecture Without Architects: How AI Coding Agents Shape Software Architecture

Phongsakon Mark Konrad, Tim Lukas Adam, Riccardo Terrenzi, Serkan Ayvaz

TL;DR

AI coding agents are making significant architectural decisions implicitly, shaping software systems without explicit developer oversight. This paper introduces "vibe architecting," where prompts dictate infrastructure, and proposes patterns and practices to govern these emergent architectural choices.

What Problem Does This Paper Address?

The rapid rise of AI coding agents, capable of scaffolding entire projects and integrating complex functionalities, presents a new challenge for software architecture. Traditionally, architectural decisions are deliberate, documented, and reviewed. However, current AI agents make these decisions (e.g., framework selection, database configuration, integration protocols) in seconds, often without justification or record. This "agent-driven architecture" leads to "vibe architecting," where the system's structure is shaped by the nuances of natural language prompts rather than intentional design. This opacity creates hidden technical debt and a widening gap between rapid AI-driven development and the slower pace of human review.

Methods

The authors employ a two-pronged approach:

  1. Tool Survey and Mechanism Identification: They surveyed six AI coding tools to identify five key mechanisms through which agents make implicit architectural decisions:
  2. Model Selection: Different LLMs produce structurally different code.
  3. Task Decomposition: How agents break down tasks influences modular structure.
  4. Default Configuration: Agents rely on training data priors when no explicit rules are set.
  5. Scaffolding and Autonomous Generation: Templates and full project generation embed architectural choices.
  6. Integration Protocols: Tools dictate how systems integrate with their platforms.
  7. Illustrative Case Study and Pattern Proposal: They conducted a case study building a customer-service chatbot using three prompt variants with Claude Code. This demonstrated how subtle prompt changes lead to structurally different systems. Based on this and analysis of frameworks like LangChain and LlamaIndex, they propose six "prompt-architecture coupling patterns" that map prompt features to required infrastructure. These patterns are categorized as either "contingent" (likely to weaken as models improve) or "fundamental" (logically necessary).

Datasets

Not discussed in the paper. The paper focuses on analyzing existing tools and a custom case study.

Key Insights and Ideas

  • "Vibe Architecting": The core concept that prompts, not deliberate design, are increasingly dictating software architecture. This shifts the focus from code generation to prompt engineering as a primary architectural activity.
  • Implicit Architectural Decisions: AI agents are making architectural choices at scale and speed, bypassing traditional review processes. These decisions are often buried in generated code, lacking documentation or rationale.
  • Prompt-Architecture Coupling: Specific features within prompts directly necessitate certain infrastructure components. For example, requesting structured JSON output requires parsers, validators, and retry logic.
  • Fundamental vs. Contingent Couplings: Some infrastructure needs (like tool execution orchestration) are fundamental and will persist, while others (like native JSON output) might become less critical as LLM capabilities advance.
  • The Speed-Review Gap: The immense speed difference between AI code generation and human architectural review is a critical challenge.

Key Results

The case study demonstrated significant architectural divergence based solely on prompt wording:

  • Variant A (FAQ): Resulted in a flat architecture with 141 lines of code (LoC) and 2 files.
  • Variant B (Structured JSON): Required a Zod schema, retry handler, and fallback generator, increasing to 472 LoC and 4 files.
  • Variant C (Tool Access): Introduced a tool registry, agent loop, and SQLite state store, leading to 827 LoC and 6 files.

This shows a 5.9x increase in LoC and a 3x increase in files for the same task, driven purely by prompt specificity. The paper also cataloged six prompt-architecture coupling patterns, illustrating how features like structured output, few-shot examples, and function calling necessitate specific infrastructure.

Limitations and Future Work

Limitations:

  • Construct Validity: The definition of "architectural decision" and "vibe architecting" are analytical constructs proposed for discussion.
  • Internal Validity: The case study used a single tool and LLM; further testing with different agents, models, and developer populations is needed to confirm generalizability. Quality impact assessments are analytical, not empirical.
  • External Validity: The tool survey reflects the landscape of early 2026, which is rapidly evolving. The proposed patterns are not exhaustive.

Future Work:

  • Cross-Agent Replication: Investigate if different agents produce similar architectural outcomes for the same prompt.
  • Architectural Footprint Metrics: Develop richer metrics beyond LoC and file count to better quantify complexity and flag potential issues.
  • Proactive Governance Tooling: Create tools that preview architectural impacts before code generation, addressing the review-speed gap.
  • Automated ADR Generation: Extract architectural decisions from agent reasoning traces to automatically generate Architecture Decision Records (ADRs).
  • Pattern Composition Analysis: Study how multiple coupling patterns interact and scale when combined in complex systems.
  • Three-Layer Framework: Develop and validate a framework mapping tool mechanisms to software architecture concepts for architecture-aware AI-assisted development.

Summary updated

BibTeX
@inproceedings{konrad2026architecture,
  author = {Phongsakon Mark Konrad and Tim Lukas Adam and Riccardo Terrenzi and Serkan Ayvaz},
  title = {{Architecture Without Architects: How AI Coding Agents Shape Software Architecture}},
  booktitle = {ICSA-C},
  year = {2026},
  doi = {10.1109/icsa-c68850.2026.00087},
}
Figure