Proposing Methods to Extend Evolutionary Architecture with AI Agents
Hemant Kumar Mahato and two colleagues have proposed a new architecture evaluation method called 'Agent-Based Fitness Functions' that combines AI agents with version-controlled evaluation criteria. The approach is presented as a mechanism for AI agents to continuously and automatically evaluate concerns that require judgment—such as design intent, boundary validity, and obsolescence of design documentation—which are difficult to capture with conventional numerical metrics.
A new evaluation methodology utilizing AI agents has been proposed within the concept of 'evolutionary architecture'—an approach for continuously assessing and maintaining software architecture quality even as systems change and grow over time. Hemant Kumar Mahato, Łukasz Sieczkowski, and Vijayasenthilkumar Kuppusamy have presented an approach called 'Agent-Based Fitness Functions'.
Evolutionary architecture is a concept for maintaining design-quality standards as systems evolve and scale. Traditionally, 'fitness functions'—the mechanisms for measuring this quality—have focused on metrics that can be clearly quantified numerically, such as test coverage and response time. However, matters that are difficult to quantify and require human judgment—such as design intent, boundary validity, and inconsistencies in responsibility allocation between modules—have proven challenging to evaluate adequately through conventional automated checks.
The approach proposed here is characterized by combining AI agents with version-controlled evaluation criteria (rubrics). Specifically, three perspectives are identified as evaluation targets: 'boundary fidelity', which indicates whether module and service boundaries are maintained as intended; 'contract drift', which indicates whether semantic contracts between services have diverged over time; and 'ADR premise obsolescence', which assesses whether architectural decision records (ADRs) remain current with the system's actual state.
Version-controlled rubrics refer to a system for maintaining evaluation criteria themselves under version control, similar to source code. This enables transparency in the evaluation rationale while allowing evaluation standards to be updated when system or organizational circumstances change. AI agents operate based on these rubrics to perform judgment-like assessments—resembling human review—continuously and automatically.
The importance of this approach stems from the increasing complexity of software systems and accelerating development velocity. In large-scale systems, it becomes difficult for humans alone to continuously track the impact individual changes have on overall design consistency. Continuous evaluation loops powered by AI agents are positioned as complementary mechanisms for early detection of such gradual quality degradation.
Conversely, evaluation by AI agents introduces challenges around ensuring evaluation accuracy and consistency of judgment. Since rubric design itself requires specialized knowledge and continuous maintenance, adoption costs cannot be considered low. Nevertheless, the direction of including difficult-to-judge design concerns within the scope of automation represents a noteworthy approach insofar as it brings new options to architecture governance.
Going forward, the extent to which such agent-based evaluation methods can achieve adequate precision and operational cost in actual development environments will be critical to their adoption. As methodologies for rubric design and management mature and integration cases with existing CI/CD pipelines accumulate, the practical utility of this approach will become increasingly clear.
This article is an original work independently written and edited by the AI issue editorial team based on factual reporting. © AI issue. Unauthorized reproduction, redistribution, or use for AI training is prohibited.