From Skepticism to Cautious Optimism - My Foray Into Coding Agents

This is my article documenting my journey into the emerging world of AI coding models. Writing code by hand has been a big part of my life for approximately 20 years at this point and has historically served as a source of great joy and creative empowerment for me.

However the rise of vibe coding is impossible to ignore, so over the last year I decided to relinquish my preconceived biases and give it a try for myself.

Early skepticism and resistance

Vibe coding, for those unfamiliar with the term, is a software development paradigm where instead of writing code for a project by hand, the developer will instead prompt an AI agent to write the code on their behalf.

As the capabilities of these coding agents advanced, so too did my bitter resentment of them.

Coding had always been my passion in life. It was a skill I had painstakingly honed over many years and indeed was something that embued me with a great sense of pride.

Yet it felt as though day by day, machines were replacing me.

In addition to feeling like my favourite thing in life was being swiftly taken from me, I had other concerns:

If this is the future, will my skills atrophy?

Will I become beholden to a small handful of behemoth companies such as Anthropic or OpenAI?

Will the innovation that coding has inspired in me over the years disappear?

For these reasons and many more still, I resisted using these tools for a long time.

So what changed?

Factors of transition

There were a couple of factors that influenced my decision to at the very least give coding agents a try:

The first and biggest reason was my discovery of open weight/open source models that can be run on local hardware via tools such as Ollama.

Although they're not particularly useful on lower spec hardware at the present, I can only anticipate this sector will grow over time. As models become further optimized and consumer hardware improves, I believe there will come a point where they will be good enough to sever our dependence on paid, hosted models.

I actually made a video experimenting with running coding models on my own M1 Pro MacBook a little while ago:

The second reason I decided to give them a try was a slightly more nuanced realization of the difference between replacing developers with AI vs augmenting developers with AI.

I realized that instead of looking at it as a complete overthrow of the role, it's more of a shift into a related role. I began to look at it through the lens of a software consultant as opposed to through the lens of a software developer.

Instead of building projects at a low level, I came to think of it as directing virtual developers by architecting technical projects and designing the systems that would achieve the intended features.

This shift in perspective made me feel a lot more comfortable with the notion of machines writing the majority of the code instead of myself personally because I'm still involved in reviewing and overseeing the technical processes and implementation strategies.

So with all that said, where does this leave me?

Cautious optimism

Until such a time that local models are a feasible option, I'm content using a paid/hosted LLM service. My hope is that RAM prices will eventually return to normality and advancements in local models will one day reach a point where these models can run on our own hardware.

A good friend of mine encouraged me to give Claude Code a try, so I did.

For a long time I've wanted to build a web VR project for my Quest 3, using Three.js for the 3D rendering. I had an idea and thought this was a good use case to test out Claude.

With it I built Orbis XR, a web VR app which in fact leverages a local small language model on a web worker for speech-to-text (more on that in a future article), then feeds that text into Google's Nano Banana 2 image generation model.

The idea of the project is you describe a scene or environment, then it will generate a spherical panorama and render it in 3D space around you in VR.

I was pleasently surprised by how much I enjoyed the workflow on this project. I found the model was quite capable of understanding the nuances and finer details of my prompts, and implemented them in a way that I was happy with.

As much as I don't love the fact that I am currently relying on a paid service by Anthropic to achieve this agentic workflow, I have faith that this style of development will one day be achievable on consumer hardware.

An additional feature that I must give Anthropic credit for is their remote control sessions. I've yearned for as long as I can remember for an "on-the-go" workflow for building projects from my phone. However between clunky UIs and typing syntax on virtual keyboards, it never really manifested until now.

So with all that said, I'm left in a state of cautious optimism.

My views on agentic coding have really shifted, and while I'm not comfortable referring to myself as a "vibe coder", I do plan on leveraging agentic workflows moving forwards.

Until next time,

Kirk M. (@saricden)