My Site

Architecture Exploration and Reflection Meet LLM-based Agents

Diaz-Pace, J. Andres, Tommasel, Antonela, Capilla, Rafael, Ramirez, Yamid E. (2025) 10.1109/ICSA-C65153.2025.00015
2025 IEEE 22nd International Conference on Software Architecture Companion (ICSA-C)
Brief Presents ReArch, a proof-of-concept LLM-based agent that adapts ReAct and LATS techniques for autonomous architecture exploration. Agents search for architectural patterns and tactics, reflect on pros and cons, and recommend ranked design decisions.
Abstract The exploration of architecture alternatives is an essential part of the architecture design process, in which designers search and assess solutions for their requirements. Although automated tools and techniques have been proposed for this process, they still face adoption challenges. Nowadays, the emergence of generative AI techniques creates an opportunity for leveraging natural language representations in architecture design, particularly through LLM-based agents. To date, these agents have been mostly focused on coding-related tasks or requirements analysis. In this work, we investigate an approach for defining design agents, which can autonomously search for architectural patterns and tactics for a particular system and requirements using a textual format. In addition to incorporating architectural knowledge, these agents can reflect on the pros and cons of the proposed decisions, enabling a feedback loop towards improving the decisions’ quality. We present a proof-of-concept called ReArch that adapts elements from the ReAct and LATS agent frameworks, and discuss initial results of applying our LLM-based agents to a case study considering different patterns.
Keywords software architecture, llm agents, design patterns, react, architecture synthesis

1. Summary

  • Generate ranked decision candidates (with explanation) for the software architecture design process via tree search.

2. Key Points

3. Handwritten Then Keyed Notes

3.1. Introduction

  • Proposes "ReArch": a design agent that incorporates architectural knowledge and explores decision spaces and reflects on proposed decision candidates to produce a ranking of design decisions for a given set of requirements within a given system context.
  • Explores decisions via 2 approaches (and an added reflection step):
    • ReAct
    • Language Agent Tree Search (LATS)
  • A previous project, "ArchMind" gave users a set of design tools which were manually invoked, this project attempts to be more autonomous by simply providing an output of ranked proposed decision candidates.
  • The goals are:
    • increase quality of analysis & decision making during design process when discussing alternatives
    • reduce barriers of formal representations, specialized technical knowledge, increase accessibility for varying levels of expertise
  • Retains manual design process while leveraging reasoning & generation capability of Large Language Model (LLM).

3.2. Background

3.2.1. Agent, LLM

  • The concept of an agent has evolved more than once. In the current form, the LLM acts as the "controller" of the agent.
  • Agents are typically composed of: 1. memory, 2. tools & 3. a decision-making mechanism. An LLM-based agent's may be then composed as follows:
    • Memory: storage for in context learning, tracking tasks
    • Tools: actions available to agent for interacting with environment
    • Decision-making mechanism: logic that orchestrates LLM calls and tools via specialized prompting technique (e.g. ReAct, LATS)

3.2.2. ReAct

  • ReAct is a method of utilizing a LLM in cycles of: reason -> act -> observe -> continue ? exceed max cycles -> stop
  • Authors suggest it can be employed in software architecture domain but must be equipped with adequate tools and instructions.
    • Challenges:
      • varying response quality
      • refining reasoning
      • exploring alternatives
    • Above can be partially addressed by adding a "reflection" step:
      • Reflection is an opportunity to store a memory
      • Reflection can:
        • assess all requirements met
        • judge decision quality
        • identify tradeoffs
      • Reflection may trigger:
        • exploration
        • additional subsequent decisions

4. Annotations

4.2. Introduction

4.2.1. Previous approaches relying on formal representations present adoption challenges.

Although automated approaches have been proposed for these tasks [1, 2], many of them rely on for- mal specifications of requirements and architectural models to leverage search methods (e.g., evolutionary algorithms, model checkers) and quality-attribute solvers. Other works have framed architecture synthesis as a planning problem. While these approaches have shown promise, their reliance on formal representations presents practical adoption challenges.

5. Acronyms

LLM Large Language Model 1, 2, 3, 4, 5

LATS Language Agent Tree Search 1, 2