It's time to talk about coding again...

Kelton Halbert

Published: 2026-08-99

Contents

It’s time to have a chat…

I’m not really one to be easily drawn into hype cycles. Maybe it’s just the old millennial hipster in me, but the more someone shouts about something the less inclined I am to take it seriously. It’s like some sort of self defense mechanism for avoiding things people put suspiciously too much effort into selling you on. I’ve found this to be both a helpful and detrimental attitude to have. I never got into the cryptocurrency hype… in fact, I’ve never so much as bought a single coin. Sure, maybe I missed a chance to make some money on BitCoin, but I also avoided losing it too. I most certainly never took NFTs seriously and saw them for the scam they were (the technology itself is whatever, what people did with it though…). So, it should be no surprise that my defenses have been raised ever since generative coding tools became available.

Like many, I tried GitHub CoPilot when it first arrived, then messed around with ChatGPT and Gemini (NOAA/Google partnership meant we had it for free). If a new tool comes out that can potentially make me better at my job, or make my job easier, of course I want to investigate it. My investigation quickly came to the following conclusion: the code quality of these tools suck, the hallucination problem resulted in overconfidently selling the wrong solution to the end user, the tools did not foster understanding well, and the claims made by the AI labs were egregiously ahead of the real-world capabilities and use cases. I was not impressed and returned to coding the old fashioned way (now being referred to as “Trad Coding”). That was 2024-2025, and the conclusions withstood periodic re-testing as updates and innovations arrived.

During that time, generative AI and its tooling became an increasingly divisive topic, and I found myself sitting squarely in the “anti” crowd. However, towards the end of 2025 and into early 2026, a lot of noise was being made in the tech space about this new paradigm they were calling “Agentic Coding”. For a long time time I scoffed, rolled my eyes, and plugged my ears. I had already tested this stuff out and concluded it was garbage. Eventually, that noise became too loud to ignore. The signal in the professional software development space was increasingly incongruent with my experiences, so I decided to do what any scientist should do.

I set aside my bias. I set aside my preconcieved conclusions. I decided that I needed to take an evidence based approach to this new “Agentic” paradigm, since it did appear quite different to my initial tests and conclusions from prior investigations. I would conduct an experiment in a space I am intimately familiar with, but I would use the tools as they were designed/intended and see how the process shakes out.

Conclusion: I had to eat my words and change my perspective and conclusions. At the risk of invoking the ire of the anti generative tools crowd, these tools have real value and I observed tangible benefits and improvements. I also observed that some of my concerns about using these tools were completely valid, and I now had to reconcile living in a world where agentic coding can be both useful and harmful.

Rough sketch of ideas/points to make…

  • Coming from a skeptical, evidence-based perspective

    • Understanding the polarizing nature of generative AI
    • Resistive of hype cycles and marketing
      • Never bought crypto/NFTs
    • Code quality in the early iterations of these tools was bad
    • THESIS: Put aside preconcieved notions and biases — embrance the tools as they are intended to be used and investigate

  • Code Quality

    • I’ve seen meteorological code, and LLMs can certainly meet/exceed that quality now
    • GRAPHIC: Normal distribution showing the percentile range of generative code vs expert code vs meteorological code
      • Inteded to address how it tends to pull from the middle ranges of the distribution, but this can elevate meteorological code
    • LLM coding may actually be one of the more appropriate uses for this technology vs other applications

  • Agentic Coding is a Paradigm Shift

    • Fundamentally different from and superior than typical chatbox based approaches
    • You are in the drivers seat, and you have immense influence on the quality of the output and design
      • As with traditional code, the quality is influenced by the expertise of the one using the tool
    • Talk about tools (Claude Code/Codex/Pi)
    • Ask questions of the tool
    • Parallel tasks, adversarial review, multi-model councils

  • A Perspective on Where Agentic Coding Fits

    • Projects where you know how to do something, but may not have the time, energy, $$, or effort available to prototype that project
    • Repetitive tasks and scripts (e.g. making plots)
    • One-off/throwaway experiments/investigations
      • fail/learn faster
    • Refactors, code analysis, edge case detection, code exploration/understanding, architecture review, performance/security review
      • Code generation becomes cheap, and if you’re willing to throw it all away and start over from scratch, you start to understand the value
      • Refactored the Heimdall render/loop pipeline from the ground up multiple times in a period of a few afternoons, would have taken weeks or months

  • Maintaining Expertise

    • Compare/contrast the age old practice of copying google/stackoverflow solutions
    • Compare/contrast importing an external library to solve your problem
    • Open-ended concerns about how to still learn without as much friction
      • Perhaps it is discipline… those who want to learn and develop expertise will, and those who copy/paste answers from SO without understanding will continue to not

  • Social, Sociatal, Economic, Ethical Considerations with Generative AI Tools

    • Intellectual property issues. The AI labs have stolen and harvested your data to sell it back to you in a monthly subscription
    • These tools require lots of compute, which requires data centers and all the fun associated with that…
      • water usage, real-estate buyouts, community impacts