Human in the Loop

The Citizen Developer Becomes the Human-in-the-Loop

• By Howard M. Cohen
• December 22, 2025

This column launched as "The Citizen Developer" on October 17, 2022. By April 2023, we announced, "Coming to a Low-Code Platform Near You: Artificial Intelligence." The change we're completing today began soon after we started.

By August, we were preparing readers for "Artificial Intelligence and the Citizen Developer," heralding monumental change yet to come.

Just two years ago this month, our end-of-year message was, "Outlook 2024: The Citizen Developers are Coming! The Citizen Developers are Coming!" In retrospect, we sound hopeful.

March of 2024, we asked whether AI would replace the Citizen Developer. Then in June, we asked, "Where Did Low Code/No Code Go?" (No apologies to Bob Dylan.)

And with that, in just three short years, we find ourselves leaving the sliding tiles of LCNC behind and vibe coding to our heart's content.

Speaking of Vibe Coding
The term "vibe coding" was first introduced by OpenAI Co-Founder and former Tesla Senior AI Director Andrej Karpathy in February 2025. We first featured Karpathy in The Citizen Developer the following August, when we asked, "Are You Ready for Vibe Coding? Yes, You Are!"

Karpathy, both a gifted speaker and a genius, has since gone back and forth on several issues related to vibe coding, like the role of context engineering and the efficacy of the term itself. He now emphasizes the need for discipline and rigorous human verification.

AI as a Savant Child
Recently, Karpathy has taken to exhorting us to treat AI as a "junior intern" rather than an autonomous expert. Speaking on this subject, he warned, "The emphasis is on keeping a very tight leash on this new over-eager junior intern savant with encyclopedic knowledge of software, but who also bullshits you all the time."

He has also observed, "Cognitively, even the most advanced models feel like 'savant kids.' They have perfect memory but don't really know what [they are doing]... They can convincingly create all kinds of slop that looks really good."

Clearly, these 'savant kids' are not his favorite children. In fact, Karpathy has been very vocal recently about the dangers posed by unsupervised agents. He warns against unleashing them too early. He clarifies that while the models possess 'perfect memory' and can pass exams to qualify for a Ph.D, they lack a fundamental understanding of what they are actually doing or the consequences of their actions.

Autonomy Doesn't Mean Unsupervised
If you come away from reading this article only remembering this headline, we will have achieved a significant step forward.

When people speak of their fear of the consequences of AI, most are referring to fully autonomous agents that you build and let go off on their own to do their job. They feel it's too easy for those agents to slip off the rails and get into dangerous territory.

They're not wrong. Software is software. Things can go wrong.

There is a continuum across which agents progress from simply assisting their human users, to a copilot status, to a more sophisticated copilot that can access and use tools, become semi-autonomous, and finally fully autonomous. The fact is that very few use cases have been identified for fully autonomous agents, and most are using simple assistive agents most of the time.

Digital Employees
In May 2025, we congratulated readers on their promotion to manager. Manager of digital employees, also known as AI agents. This is a healthy model to strategize around, as all employees require some level of supervision, including these new digital employees. Human oversight is core to success.

We are already seeing the emergence of AI-specific protocols such as MCP and A2A that enable multiple agents to work together. Especially in the early days of this multi-agent environment, misalignment of mission is all too possible. Then the agents can end up reinforcing each other's mistakes. Only a human operator can catch these problems.

Security Risks
We are only now starting to see providers emerging to protect our AI investments with adequate security, governance, and compliance monitoring. AI employs mammoth volumes of data, making any exposure a potential disaster simply because of the sheer volume of compromised data. Any security expert will tell you that one of the most essential elements of their work is having the intuition to identify subtle signs of impending danger. Only human supervisors have that intuition. It's among the qualities nobody has yet figured out how to program into an LLM.

Take Aways
These suggested takeaways from this article set the stage for everything else we're going to be covering here in our newly renamed "Human-in-the-Loop."

• Stop fearing fully autonomous AI agents. Almost nobody is building them, and soon everyone will agree they require as much supervision as any other employee.
• Start to recognize the 'unprogrammable' when you see it, the things humans can do that digital employees or agents will never be able to do. These include intuition, emotional response, judgment, unreasonable expectations that nonetheless get fulfilled, and the mysterious history lurking in our DNA and informing so much of why we do what we do the way we do it.
• Drop the word "autonomous from your vocabulary and think of AI Assistants instead of AI Agents.
• Think of everything you'd personally like to be able to do faster and more thoroughly. Then start building a prompt for the model of your choice to do that for you. Keep discussing it until it's giving you back hours or days of your time.
• Remember that the things they can do, we can do; they do them far faster, and they have ready, immediate access to far more information than we can.
• Consider your AI model to be that savant child who can earn a Ph.D but has no clue about what they're doing with it.
• That child needs you – it's human-in-the-loop!