Reasons behind the widespread disappointment in AI’s capabilities

Understanding the Reality and Future of AI Automation in Software Development

In recent times, public sentiment towards artificial intelligence has largely been lukewarm, and it’s easy to see why. Much of the mainstream discussion is dominated by commercial interests, with tech entrepreneurs promoting quick-fix solutions—often involving superficial tools or dubious claims—that fail to address genuine needs. The landscape can feel like a chaotic marketplace packed with empty promises and unnecessary hype, often designed to profit rather than genuinely innovate.

However, beneath this noisy surface, there are quieter, more meaningful developments driven by programmers and developers. Many are leveraging AI not as a shiny new gadget, but as a powerful tool to create custom automation processes that streamline workflows and improve efficiency. These implementations tend to be highly specific, tailored to individual needs, and are not always easily transferable between different systems or projects. Achieving a comprehensive, all-in-one AI solution remains a distant goal, and likely will for some time.

One important insight is that the true potential of AI isn’t about automating entire jobs outright. Instead, the real breakthrough lies in automating the process of automation itself. Typically, automating a single task involves a series of steps—often five or fewer—some of which may involve looping or memory systems, and interactions with various APIs.

At first glance, this might seem straightforward: define prompts, sequence steps, store and retrieve data, connect to external services. But each step requires careful attention—crafting precise prompts, structuring memory, and ensuring correct order—making the process more complex than it appears. This complexity necessitates multiple intelligent agents: one to generate prompts, another to design the overall architecture—including memory management—and a third to handle API interactions.

The good news is that many of these capabilities already exist. AI systems have been generating their own prompts for quite some time. For example, recent research (see this 2023 paper) demonstrates how structured protocols like MCP can embed instructions within API calls, enabling more autonomous operation. Additionally, tools like AgentForge now incorporate YAML-based architectures, allowing AI to independently assemble and manage sophisticated workflows—sequencing prompts, integrating memory, and handling API calls—all without manual coding.

What does this mean for the future? We are approaching a point where most of the tasks involved in setting up and managing AI-driven automation can be fully automated. This isn’t just a fleeting trend; it is likely the final frontier in automating repetitive