# My Journey with AI So Far (June 2026) > Source: > Published: 2026-06-30 14:21:06+00:00 When early AI first appeared in the tech scene, like many others I was skeptical of it. I kept an eye on it from a distance, but ultimately didn't invest too much time into it given its abilities seemed a bit gimmicky at best. Since then, a lot has changed in the industry and I wanted to take a moment to share my journey with AI thus far and capture a snapshot of where I am with my AI workflow today. ## My first step with "AI" My first "real" encounter with AI started back when Copilot was providing "code predictions." If you haven't used it, imagine if instead of [IntelliSense](https://code.visualstudio.com/docs/editing/intellisense?ref=bencodezen.io) where it autocompletes text derived solely from your codebase, VS Code tries to anticipate what you want to type as you write the code. In this example, you'll see that VS Code is trying to anticipate what goes in the function named `calculateDaysBetweenDates` with the parameters of `begin` and `end` . While it was fun to play with, I found it often interfered with my ability to think through a problem because it was always jumping ahead to try and come up with a solution. And though it could be helpful at times, ultimately I found it didn't contribute more than it distracted me. ## Enter ChatGPT After that phase had come and gone, this was around the time that AI chatbots like [ChatGPT](https://chatgpt.com/?ref=bencodezen.io) really started to make some noise. People were asking it for all kinds of things and seemed to be getting results ranging from "Holy smokes it one shot the app I wanted it to build" (which was clearly exaggerated) to "Why did it recommend using glue for cheese on pizza?" The biggest obstacle I found with this phase of AI was the fact that people were impressed by the occasional right answer even though they had to wade through a lot of hallucinated answers at the same time. So while it was interesting as an experiment to see if it could figure out things I was stuck on here and there, ultimately it wasn't reliable enough to be a big part of my workflow. ## Hello Claude Code! The real breakthrough for me was the more integrated experience of [Claude Code](https://claude.com/product/claude-code?ref=bencodezen.io) in my terminal. For the first time, I was able to actually work iteratively while using version control with Git to ensure that I could understand what was happening. Even though there were still issues like hallucinations and false leads, the workflow felt more like companionship than me simply asking a question and hoping for the correct answer. This would also lead me to experiment more with VS Code's Copilot for a more integrated workflow. While I still love it for working on a single project, I found it to be limiting when working on concurrent projects since I struggled to get proper notifications working with VS Code when agents would finish their work. In addition, it became rather tricky to manage all the different workspaces which made context switching even more expensive. ## More Claude Code is Better? So, naturally, like any engineer looking to increase their effectiveness, I wanted to explore the possibilities of running parallel agents across projects to make myself "more efficient." And one of the workflows I'd establish would use [cmux](https://cmux.com/?ref=bencodezen.io) to keep multiple terminal tabs running with notifications to ensure I knew when certain agents were done with their tasks. While it felt very productive, I hesitate to say that the outcomes from this workflow were as impactful as I had hoped. There is certainly a lot going on at any given time, but the harsh reality is that the bottleneck of reviewing and having enough time to deeply think about a problem makes this type of work unsustainable. ## Where am I now? As of this point, I can confidently say that the current state of frontier agents has become a big part of my workflow because I've found ways to minimize their hallucinations and stay focused on the problem at hand. **Model:** I use both Anthropic's Claude Opus 4.8 (1M Context) and OpenAI's ChatGPT-5.5 as my daily drivers to compare responses and review the work being produced. If I had to choose one, I'd lean toward Opus 4.8 (1M Context) as my go-to, and would use Fable 5 for more complex tasks if I had access to it.**Harness**: As far as parallelization goes, I haven't given up on it yet. I'm currently experimenting with [Conductor](https://www.conductor.build/?ref=bencodezen.io), which allows you to manage multiple agents in a more efficient manner than a terminal. Early results seem positive! That said, I'm still working out the nuances on how the relationship with AI changes when parallelization is introduced. In other words, what constitutes a healthy workflow vs appearing busy while having little real world impact. I believe there's a place where it can be used to enhance output, but more on that as I develop a better theory. What I can say with conviction, though, is that the biggest mistake I've found with working with AI is treating it as a separate entity that you order around. When approached in this fashion, it almost always ends up with the user offloading their critical thinking and blindly trusting the AI's output. Rather than seeing AI as "Artificial Intelligence," I think it is best to think of it as "Augmented Intelligence" instead. And when I say "Augmented Intelligence," I am referring to augmenting your own intelligence, not the model itself. Because while there is a lot of chatter about how to improve your agent's efficacy, at the end of the day, I believe that you should walk away from each working session with AI feeling more capable and knowledgeable. If you can find a workflow that lets you do that, you're on the right path to finding a sustainable workflow for yourself.