For the past couple months, I really delved into AI Agents for developing applications. My journey got more serious when my company rolled out a program for AI Acceleration by giving the Product & Tech division a $50 stipend for subscribing to an AI Agent of your choice. It was a great introduction, especially when these tools become more and more capable at helping write code.
As time passed by, I started to offload more and more tasks to the AI Agent. My tool of choice is Augment Code, because I heard a lot of great feedback about it. It’s also the easiest tool for me to adopt as it basically lives on VS Code, which is suitable with my existing workflow. It was great and boosted my productivity, helped reduce mistakes, and I could even publish a full-blown full stack application to help my team keep track of our terraform staleness.
Personal Project
Learning from day-to-day AI experience, I started to subscribe for my own usage. I started with Cursor to learn the tool, as many of my colleagues use it. I needed to adjust my workflow a little bit, but turns out it’s not that bad, moving from VS Code to Cursor.
But for the past couple weeks, I heard great feedback about OpenCode. Many people on Twitter shared their experiences using OpenCode combined with the Claude Opus model. People even started showing their billing that reached $200 per month for their development. I was curious and traded my Cursor subscription for Claude, with initial $20 credits. Then it grew to another $20, and another $10.
Why? Because it’s that good! The model can really help me develop my own project quickly, with great accuracy. I can focus more on something I enjoy, like thinking about the product, doing CI/CD, and deployment, etc. The model really helped me roll things out until I kinda ran out of ideas.
With this setup, I was finally able to publish my old ideas like recreating a successor of Teller named Sstart, a dead simple secret manager called Taplock, and even my very old project to sync personal calendar events to work calendar to ensure there is no meeting scheduled at the same time Busical. All the projects were developed and published in the same month! I can’t even imagine I could do that last year, because once I tried to start it, all other non-functional stuff came in: choosing platform to deploy, which framework to use, UI library, and more. Within several days, I already abandoned it, because it felt not fun anymore.
What did I learn
From that journey, I learned several things:
Do planning before building
This is the biggest point I learned, especially when you pay your own subscription. Using the expensive model for immediate building produces back-and-forth feedback and burns a lot of tokens (that’s why I bought credits that high). Once I did more planning before building, with all the clear todos that match your intention, the build will go smoothly and your spend goes to the correct implementation.
With that, I can decide where to spend the tokens that suits my needs. If I know what to do, I can spend on cheaper models for planning, because I kinda know my expectations, and spend more on implementation to ensure the model implements it correctly, including some details like code patterns, etc. And do the other way around when I need to brainstorm, do refactoring, or introduce new approaches, needing a “smart” model to do it better.
Software Development is changing fast
This AI setup and all really changed me, on exploring ideas and actually delivering. In my opinion, Software Engineers will be changed a lot, writing code is not the same anymore. It’s starting to become less and less important, more people can start creating their own products, seeing them developed and published, giving value to others without needing to have deep understanding of Software Development. I think I finally really started enjoying product development again.
Key Takeaways
I would highly recommend you folks to try it. Try to utilize it, even with the free models. The landscape is changing, and we need to adopt. Fundamentals are becoming more and more important, as we can write code faster and ensure the implementation is correct with less to no issues.