AI-Driven Development in Scholarly Publishing

The past 18 months have been a whirlwind in the world of technology and scholarly communication. Since AI agent-based development became widely accessible via the web, progress has been astonishing—moving from simple code suggestion tools to autonomous agents capable of developing entire applications. Most of this transformation has taken place in just the last six months.

Insights from the Frontlines

I've been working with AI agents in application development for the past 18 months, but intensively in the last few, and I’ve learned a lot. Here are some key insights from my hands-on experience:

Human-in-the-Loop is Essential

AI primarily understands applications as code. It relies heavily on human input to refine features, verify functionality, and troubleshoot unexpected behavior. Currently, if an application fails or behaves unpredictably, a developer must step in to diagnose the issue. However, this will likely change with the rise of "agent operators"—tools that use computer vision and browser automation to navigate web interfaces autonomously. Once these tools mature, AI agents will gain a visual understanding of front-end issues, enabling them to diagnose and resolve problems more effectively.

Code Clean-Up is Frequently Overlooked

AI can rapidly build applications, but it doesn’t inherently optimize or clean up its own code. Without explicit instructions to tidy the codebase, AI risks iterating over outdated or inefficient structures, leading to wasted time and resources. Unless addressed—either manually or through automated processes—this oversight can hinder long-term development.

AI Agent Context is a Limitation

One of the biggest challenges when developing applications with AI agents is their limited context window. Over time, the agent "forgets" key aspects of the application—why it was built, how it works, and even what it was working on. This is somewhat like working with a highly forgetful developer who needs frequent reminders about the project. As a result, the AI may generate redundant or inefficient code, solving problems without utilizing existing solutions already built into the application. This can lead to code bloat, requiring additional cleanup and oversight to maintain efficiency. While not a dealbreaker, it adds friction to the development process and highlights the need for better memory retention in AI-driven workflows.

Quality Feedback Enhances AI Efficacy

Providing structured feedback is not just beneficial—it’s necessary. AI agents respond best to detailed logs, browser console insights, and screenshots, all of which help refine their output. Unlike junior developers, AI agents work at scale, akin to having ten skilled junior engineers tackling a project simultaneously. This dramatically accelerates development, but effective collaboration still depends on the user’s ability to interpret issues and guide improvements. The more technical knowledge you have, the more effectively you can leverage AI in application development.

The Implications for Scholarly Publishing

AI-driven development is poised to reshape how platforms are built in scholarly communications. Large, monolithic publishing systems like ScholarOne were designed to accommodate a vast range of needs within a single framework. However, with AI, we can now develop bespoke platforms quickly—tailored precisely to an organization’s requirements—eliminating the need for complex, overly configurable systems.

So far, I’ve been able to develop some impressive platforms, albeit with a limited scope. One example is a custom CMS I built for a project focused on publishing reviews. The process was fascinating—I started with wireframes and templates, then "backwards-engineered" the use case. Drawing from my experience with workflows, I pieced together briefs for the AI to work against, allowing it to develop a functional solution. The result is, in my opinion, a very good system.

One thing that stands out in this process is the visceral experience of using a truly bespoke solution. The difference is immediate—you sense it as soon as you interact with the platform. Unlike one-size-fits-all systems, which often feel bloated due to their need to accommodate many use cases, a tailored platform is clean, elegant, and intuitive. It does exactly what you need, nothing more, nothing less. The contrast is striking, and it reinforces the idea that complexity in traditional platforms is often a byproduct of their broad, generic design rather than an inherent necessity.

To explore this concept further, I recently invited organizations on LinkedIn to collaborate on building custom platforms using AI agents. The response has been incredibly positive, leading to several real-world implementations now undergoing testing. The goal is clear: to demonstrate that agile, highly customized platforms are not only possible but superior to one-size-fits-all solutions.

Looking Ahead

AI-driven development presents both opportunities and challenges. While traditional developers aren’t being replaced, the acceleration and customization potential these tools offer is undeniably transformative. As this space evolves, I look forward to collaborating with others interested in exploring this approach.

If you have an idea or need for a platform, let’s take the plunge into this new frontier together. The shift toward personalized, efficient publishing solutions is already underway, and the future looks promising.