In the rapidly evolving field of software engineering, the intersection of artificial intelligence (AI) and human developers is becoming increasingly significant. Martin Fowler’s recent blog post sheds light on the essential role that professionals will continue to play in AI-assisted software development. Kief Morris contributes valuable insights, stating that while AI is transforming workflows, developers are not poised to become entirely “out of the loop”. Instead, many teams are likely to progress towards a model where they work “on the loop”. This means they will focus on designing the specifications, tests, and feedback mechanisms that guide AI agents rather than scrutinizing every line of generated code.
Source: MartinFowler.com
Morris categorizes the interaction between humans and AI systems into three distinct frameworks: “in the loop”, “out of the loop”, and “on the loop”. The traditional approach, “in the loop”, requires developers to meticulously review each AI output. Conversely, the “out of the loop” model allows AI systems to operate autonomously with minimal human oversight. The innovative “on the loop” model proposes that human developers should prioritize designing and maintaining the frameworks that guide AI functions, shifting the focus from constant oversight of code to creating robust testing frameworks and evaluations. This shift can enable advanced AI tools to function more effectively while minimizing technical debt.
As organizations experiment with AI-driven coding agents, a fascinating architecture termed the “agent loop” has emerged. OpenAI delves deep into its Codex system, outlining how user interactions coordinate with the model and external tools. Unlike simple command responses, this architecture allows AI to produce or modify code directly on the software environment, embracing an iterative process enriched by user feedback.
Source: OpenAI
Yet, amid the excitement surrounding AI-generated code, developer sentiments remain mixed. Conversations on forums like Stack Overflow reveal that while 84% of developers express an interest in using AI tools, trust in the integrity of AI-generated output is lacking. Often, developers find themselves expending extra effort on debugging and validating AI-generated code, reinforcing the notion that productivity gains sometimes come at the cost of increased technical debt.
The 2025 Developer Survey underscores this reality. Although a significant percentage of developers are engaging with AI tools, many are skeptical about the reliability of the code produced by these systems. The complexity of debugging AI-generated artifacts often necessitates an additional layer of review, posing challenges to the seamless integration of AI outputs into existing systems.
To tackle these challenges, some engineering teams are innovating verification and control mechanisms. For instance, Datadog has advocated for a more scalable approach to AI-generated artifacts by emphasizing automated verification. Instead of relying solely on human review, which can be inefficient, teams are encouraged to invest in automated pipelines that integrate specifications, simulation testing, bounded verification, and runtime telemetry to ensure the reliability of AI behaviors.
Source: DataDog
Datadog outlines a “harness-first” strategy aimed at agent development, which prioritizes automated verification systems. By establishing robust frameworks for assessing agent behavior through specifications and simulation testing, teams can create a more efficient environment for validating AI outputs without extensive manual inspections. This vision highlights the shift towards more sophisticated engineering practices that accommodate AI capabilities while ensuring code quality.
The overarching theme emerging from these discussions centers on the intricate systems surrounding AI development tools. As coding agents grow in sophistication, developers are urged to prioritize testing harnesses, evaluation frameworks, and observability systems that facilitate the monitoring of AI outputs. Morris’ assertion of working “on the loop” exemplifies this evolving paradigm. Rather than seeking to exclude human expertise from the development process, organizations are increasingly recognizing the value of having developers craft the guardrails that delineate how autonomous software systems function effectively.
Inspired by: Source
