Almost two years ago, one of my clients asked me to evaluate AI development tools.
At the time, I had already worked on projects that integrated AI. I had built systems that called AI services, processed AI outputs, and incorporated AI features into applications. But I had never really developed alongside an AI agent.
So I spent time experimenting.
I tried AI-enabled IDEs. I used multiple coding agents. I built demo projects. I intentionally explored different workflows to understand where these tools helped and where they struggled.
What started as a client evaluation turned into nearly two years of daily experience working with AI as part of the software development process.
What surprised me most wasn't the technology.
It was how familiar it felt.
Over the years I've spent a lot of time teaching youth through Scouting America, serving as a merit badge counselor, and helping young people learn technical and practical skills. Again and again, I found myself noticing parallels between mentoring youth and working with AI agents.
The comparison isn't perfect, but it is surprisingly useful.
AI Is Often Like an Eager Beginner
One of the biggest misconceptions about AI is that it behaves like an experienced professional.
In reality, most AI agents behave much more like a highly motivated beginner.
They have read an enormous amount of information.
They can explain concepts fluently.
They often sound extremely confident.
But confidence and understanding are not the same thing.
Anyone who has taught youth has seen this before.
A Scout may be able to repeat information from a handbook perfectly while still lacking practical experience. They know the words, but not yet the judgment that comes from doing the work.
AI often behaves the same way.
It can produce impressive-looking code. It can explain architectural patterns. It can discuss best practices.
But it still benefits enormously from guidance, review, correction, and context.
Instructions Matter More Than Intelligence
When teaching young people, vague instructions usually produce vague results.
"Go clean the campsite" and "Pick up all trash, stack the firewood, and sweep the pavilion" produce very different outcomes.
The same thing happens with AI.
The quality of the result is often directly related to the quality of the instruction.
Developers sometimes describe this as prompt engineering, but to me it feels much closer to mentoring.
Clear expectations matter.
Concrete examples matter.
Defined success criteria matter.
The better the guidance, the better the outcome.
Feedback Loops Are Everything
One of the most effective teaching techniques is immediate feedback.
A student attempts a task.
The mentor reviews it.
Corrections are made.
The student tries again.
AI development works remarkably similarly.
The most productive sessions are rarely a single prompt followed by perfect output.
Instead, they are conversations.
Review.
Correction.
Refinement.
Iteration.
The agent improves because the feedback improves.
Trust Must Be Earned
When teaching youth, you don't immediately hand someone a chainsaw and walk away.
You start with smaller tasks.
You verify understanding.
You gradually increase responsibility.
I've found that AI works best the same way.
I don't blindly accept generated code.
I don't assume correctness because something compiles.
Instead, I establish trust through verification.
Some tasks can be delegated almost completely.
Others require careful review.
The key is understanding where the agent has demonstrated competence and where it still needs supervision.
Context Is More Important Than Capability
One lesson I've learned repeatedly is that AI agents often fail not because they lack capability, but because they lack context.
This is also true when teaching people.
A student who doesn't understand the goal of a project will often make poor decisions despite having the technical skills to complete the task.
AI behaves similarly.
The more context an agent has about the project, architecture, constraints, and goals, the better decisions it tends to make.
This is one reason I've become increasingly interested in AI-friendly software architecture.
Good structure doesn't just help humans.
It helps AI as well.
The Best Results Come From Partnership
The most important lesson from the last two years is that AI works best as a collaborator rather than a replacement.
The strongest outcomes happen when human judgment and AI assistance complement each other.
The human provides direction, priorities, experience, and accountability.
The AI provides speed, recall, exploration, and execution assistance.
Neither is sufficient by itself.
Together they can often accomplish more than either could alone.
Final Thoughts
After nearly two years developing alongside AI agents, I don't see them as magical or mysterious.
I also don't see them as replacements for human expertise.
Instead, I increasingly view them as something familiar: capable assistants that benefit from guidance, mentorship, feedback, and structure.
Perhaps that's why the transition felt so natural.
The same skills that help teach young people—patience, communication, clear expectations, constructive feedback, and good leadership—turn out to be surprisingly valuable when working with AI.
The technology is new.
The human skills are not.
