Working with AI is more Mindset than Skill

Deborah Lupton / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/

A week ago, a massive ice storm hit north Mississippi and caused us to lose power. We evacuated a few days ago and are just now trying to get caught back up with work and make sense of things.

I’m trying to understand how AI agents work, and more pointedly, how education is supposed to deal with this yet another wave of AI making its way into our classrooms via Edtech providers. Grammarly has already upgraded its interface to offer a handful of simple agents that can grade, offer feedback, etc. But that’s pretty tame compared to how people are now using AI agents to code and create multiple workflows.

This is all new to me. My background is in creative writing, not coding. I read novels and short stories throughout graduate school, so I have little concept of how to code. But with agentic AI, I don’t supposedly need to. I can use storytelling or task an AI agent to code something for me that is both practical and usable in mere moments. That’s hard for me to wrap my brain around. But it’s here, and we’re once again trying to figure out how to grapple with AI, and absolutely no one has figured out agents.

Case in point, I downloaded Claude’s newish Cowork feature for my Mac and tasked the AI coding agent with turning the 10 provocative scenarios I wrote about AI and grading in a recent post into interactive digital dilemmas about AI and grading. The AI did a pretty decent job! It even walked me through how to host the activity for free. You can click on the link or image below to interact with it. When Claude needed direction, it asked me questions and guided me in shaping how I thought the output should be. In that way, the process was strikingly iterative and not automated. That alone was really something. To have AI help me refine my thinking about a project wasn’t something I expected or was ready for when using an agentic tool.

Welcome to Gas Town

I’ve read Maggie Appleton’s musings about AI for years now. She’s one of the most thoughtful voices in the coding world who is trying to figure out where AI fits into our lives. She approaches AI with a sense of skepticism and exploration, writing about both the great promise AI has to help us navigate information and likewise rob us of the things we value, like critical thinking. So when she wrote about this bizarre experiment from a fellow coder called Gas Town, I found it both riveting and incredibly challenging to wrap my head around.

Gas Town is a project by Steve Yegge. It is weird. It is bizarre. It’s as if Hunter S. Thompson started using AI and made gonzo work patterns a thing. As Appleton describes it, Gas Town is Yegge’s “Mad-Max-Slow-Horses-Waterworld-etc-themed agent orchestrator that runs dozens of coding agents simultaneously in a metaphorical town of automated activity. Gas Town is entirely vibecoded, hastily designed with off-the-cuff solutions, and inefficiently burning through thousands of dollars a month in API costs.”

Each agent gets assigned a persona and works under a hierarchy. But this is AI, so nothing quite works as it should. The agents argue with each other, ignore one another, and general chaos ensues, but what Appleton notes within this chaos are clear patterns emerging for how human beings are thinking about agentic systems.

Appleton created this graphic to illustrate some of the interactions Yegge created. It is helpful to see just how the ‘town’ is set up, and the relationships between the agents function, but the real interesting thing is how Yegge interacts and views the agents.

As Appleton notes, Gas Town isn’t a means to an end. It is a provocative piece of design fiction, but what it reveals is how human beings work with semi-autonomous machines. It also reveals where it fails. Agents don’t have judgment or taste. A human being must bring those to the table and that’s what becomes meaningful through these interactions. The human has to start asking themselves metacognitive questions that focus on thinking about how they think when using these tools:

What are the highest priority features to tackle? Which piece of this should we build first? When do we need to make that decision? What’s the next logical, incremental step we need to make progress here? These are the kind of decisions that agents cannot make for you. They require your human context, taste, preferences, and vision.

Changing How We Think About AI

Maha Bali recently completed a peer review of a preprint paper that I highly recommend reading. I also recommend reading the paper itself: Rethink Your Mental Model in the Age of Generative AI: A Triadic Framework for Human-AI Collaboration. What makes the paper and Bali’s analysis so thoughtful is the shift in thinking about working with AI as a skill you can learn or be imparted, to a mental mode of thinking about our relationship with machine intelligence and how it impacts us. Bali’s three takeaways from the article sum up the approach to creating a critical awareness of AI’s capabilities by balancing skepticism with pragmatism:

1. We need to stop treating LLMs like previous tools and recognize that what they offer is a “jagged intelligence” that is more uncertain and unpredictable. This is an important first step necessary to work with LLMs in a productive manner. Models also change all the time, and what was possible before might not be possible in the future, etc.

2. Our own mental models can get in the way of our skepticism about LLM outputs – things like the empathetic and confident tone of LLMs can make people trust it more; the natural tendency to offload cognitive tasks means it is a huge metacognitive (and even precognitive) effort to remind ourselves to constantly be skeptical and let our critical thinking kick in. Even experts may let down their guard.

3. We learn a lot about using LLMs well from practice and experience rather than being given instructions on things like good prompting. These practices are evolving and some things we thought were a good idea (e.g. “tell the LLM which role to play”) turn out not to be useful for content, just useful for form; and some other things many of us do can make LLMs perform worse (very long strings increase chances of inaccuracies getting worse and worse).

The article offers certain proposals to help us make better decisions and take steps to ensure we don’t fall into traps of overtrusting AI for important things (it’s only talking about regular use of LLMs not for mission-critical high risk uses). The proposals all sound like good ideas to teach students and anyone else, honestly!

I think these three takeaways form a solid foundation for looking at AI and the many features that we’ve seen. But this is a mode of thinking, not necessarily something that you can easily teach. That’s hard. It might be the greatest challenge posed by AI in education.

Where We’ve Been, Where We Are, & Where We’re Headed

It has been impossible to create a stable curriculum around generative AI for students or even one geared for faculty professional development. The money being poured into AI developers means nothing is stable for very long. Models change, capabilities increase in some areas, while spectacularly failing in others. The past three years have seen skill-based learning about generative AI arise and disappear with some alarming speed!

2023: Remember When We Were All Supposed to Become Prompt Engineers?

Remember that stretch of time in early 2023, when the hot new skill companies were looking for was prompt engineering? I recall seeing dozens of six-figure salaried job postings for prompt writing. Some universities even hastily set up classes or programs for prompt engineering, thinking this is the skill of the future. If you look today, you’ll notice there are no prompt engineer positions. They’re gone. AI updates in reasoning models (actually intermediate tokenization, but ‘reasoning’ sounds so much more marketable) meant you could give a chatbot a poorly written prompt and it would use its reasoning to improve it.

2024: It’s All About Getting AI to Talk to Your Data

After prompting fizzled as an in-demand skill, the next big thing was using AI to talk with your data. Big corporations spent millions of dollars designing custom RAG applications to use AI to analyze their data and create silos where outside LLMs couldn’t reach. The hot skill was analyzing outputs for errors and using AI to help navigate massive amounts of data. Except it wasn’t. Some companies are still embarking on customized solutions, but by the end of 2024, AI development whizzed past RAG applications to AI agents that used tools and reasoning models to automate workflows.

2025: Agents Galore

Agentic workflows crept into company meetings in bizarre marketing-speak that no one really understood. Simple single-agent automations or more complex Deep Research tasks that used agents to search and create full reports with hundreds of integrated sources were a dramatic departure from customized RAG workflows. Suddenly, you didn’t need AI to talk with your data; rather, you used AI to automate the entire research and drafting of reports. But then people found out that you could use agents to automate dozens of different workflows, including tasking an agent to complete a full course for you. Talk of agents saturated commercials and industry conventions. The hot new skill we were supposed to teach students was project management . . . until it wasn’t.

2026: Coding Agents Arrive

Toward the end of 2025, agents suddenly became viable for coding. You could ask AI in natural language to create a website for you and it worked pretty effectively. Claude 4.5 and Gemini 3 and newer models from OpenAI completely automated the writing of code for many people. Social media was filled with AI developers in a deepening existential crisis. Many of them were using AI to code and weren’t even looking at the output agents were generating, so what was their purpose? There was no more prompt engineering, no more verifying results, and increasingly little project management; there was simply a series of automated workflows that didn’t require much of any input from a user. How absurd has it gotten? Someone on X comically created an analog rubber stamp to digitally approve agentic workflows.

The astonishing rate of change is too volatile, too uneven, and simply too disruptive for traditional curricula to capture. That’s because working with AI isn’t a skill but a mindset. Take as many online courses about AI as you like. Most will present information that is outdated before you get your certificate. The very nature of how these tools are iterated in public transcends the notion of a standardized set of skills that can be handed down from teacher to student or training program to worker.

If we simply hand students a digital rubber stamp to approve agentic workflows, we are teaching them to sleepwalk through life without exercising the agency that’s vital for collaborating with these tools. This is why rethinking our mental models is so vital. We don’t need to teach the specific features of 2026’s AI tools—those will likely be gone by 2027. We need to teach skepticism and critical thinking as philosophical responses to machine intelligence. Instead of just changing our assignments, we must do something more provocative and arguably far more challenging. We must reflect on how we think and find value in thinking in an age where a machine is happy to do just that for us.