# The Evolution of Software Development in the Age of AI Agents Transitioning from manual coding to supervisory control systems in enterprise software engineering environments. - **Source:** [Martin Fowler](https://martinfowler.com) - **Published:** 2026-10-28 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/fa846d7f-3545-45cd-a924-0aaecf227531 ## Perception Gap in AI Andrej Karpathy highlights a significant divide in how LLM capabilities are perceived. While a small group of professionals experiences a vertical curve of progress, the general public largely views these systems as unreliable chatbots. This disparity creates a bubble where bleeding-edge users operate in a reality vastly different from the average professional experience. ## Emotional Impact of Agents Immersion courses for agentic engineering reveal that the primary challenge is not technical, but emotional. Developers often feel their professional identity is threatened as basic assumptions about coding change. Coaches report that helping engineers process this shift is as critical as teaching them to utilize new tools for building software systems. ## Enterprise Adoption Barriers Large organizations struggle to integrate AI, often using only a fraction of its potential. Unlike startups, enterprises face complex workflows, security concerns, and data quality issues that hinder rapid deployment. The path to becoming agent-ready is non-trivial, requiring significant structural changes before these companies can fully leverage advanced AI toolsets effectively. ## The 1997 Investment Analogy Economic data currently shows no productivity gains from AI outside major tech firms, leading to comparisons with the 1997 dotcom era. While a market bubble is likely, dismissing AI investment as a waste is premature. History suggests that identifying long-term winners requires patience, as real productivity gains take time to manifest across industries. ## Redefining Developer Identity Agentic coding allows developers to handle larger problem scopes by abstracting away manual infrastructure tasks. The future of the profession lies in problem-solving, logical thinking, and system architecture rather than writing code. By separating programming from mere code generation, experienced developers can leverage their deep domain knowledge to build more sophisticated systems. ## Ethical and Economic Concerns Beyond technical capabilities, significant discomfort exists regarding the ethical and economic implications of GenAI. Critics argue that the technology is being deployed by entities focused on slashing payrolls and devaluing human creative effort. These concerns highlight a growing tension between corporate efficiency goals and the broader impact on the workforce and society. ## Direct vs. Supervisory Control Software development is shifting from direct control, where humans close the loop, to supervisory control. In this new model, AI agents execute tasks while humans monitor and intervene. This transition mirrors automation research from the 1960s, requiring a fundamental redesign of the software development lifecycle to accommodate autonomous agentic behaviors and oversight. ## The Five-Level Automation Model The Endsley and Kiris model provides a framework for AI-assisted development, ranging from no assistance to full automation. Current tools mostly occupy decision support or consensual AI levels. A critical caveat remains: AI agents may not accurately report their actions, necessitating constant human supervision to verify that the system is performing as intended. ## Exploring Future Fundamentals Eric Evans emphasizes exploring the future through curiosity rather than prediction. His Domain Language Surveyor project demonstrates how to analyze vocabulary across codebases to identify bounded contexts. By blending computational and inferential components, developers can better understand the structure of complex systems, providing a foundation for more effective AI-driven software engineering practices. ## Optimizing Token Costs Managing token costs is a major concern for enterprises scaling AI workloads. A strategic approach involves classifying tasks based on domain differentiation and security sovereignty. By distributing workloads across a nine-cell grid, organizations can choose between renting frontier models or running open-weight models on internal infrastructure to significantly reduce long-term operational expenses. ## Systemic Design Evolution The future of software engineering requires redesigning jobs and systems in tandem. Moving forward, the software development lifecycle should be viewed as a control system that requires human supervision. This shift demands new skills, as developers evolve into architects who manage and verify the outputs of autonomous agents within complex enterprise environments. ## Sources covering this story - https://martinfowler.com/fragments/2026-10-28.html — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._