With how many screw-up and changes we all saw in the software development world thanks to AI agents, some are good, mostly bad to worse, I started to wonder if these AI tools are used in the real world and how they are used. So, I've been in contact with many associates I met along my career and asked them about how they approached incorporating coding agents into their development workflow. There are a few main camps:
- Pair programming mode: This is how the majority uses these tools. They run the agent side by side in the IDE. Whenever they are stuck on something, they either ask (or prompt) AI to help fix it. Through this way, people can keep up with the changes they made while minimizing cognitive atrophy. AI makes a great debugging buddy, who knew?
- Autonomous mode: Coding agents are used to generate a plan based on a feature, an app idea that you give. You then review the plan and approve it. The agent then send many subagents to parallelize the coding tasks, before returning the results to the main agent to report what they have done. This is what "the Internet" or "Reddit" or "LinkedIn" would have you believe is the only correct way to use AI, or else you are falling behind.
I have used both in various capacity. I used Autonomous mode whenever I want a major change, a feature implemented, or scaffording a new project. The problem is that these coding agents go off the rails very fast. Soon, you will have no idea what the F they just coded. And you will also inherit either cognitive debt or a massive PR you need to review for weeks. And trust me, I tried various ways to control them, even implementing another subagent to review what was coded for performance and security. Out of 5 bugs it caught, only 1 of them was of any use for me. I much prefer the other way: Pair programming mode. It is perfect for my tempo and everyone else's.
And you can trust me with these findings. I have friends in various places: in major corporations, in open source projects, in OSU departments themselves. If you keep up with LinkedIn, Reddit, or any major site, you tend to lose a grip to how the real world functions. Today, I got to visit the OSU Engineering Expo and was very surprised and excited at many showcases. Like, these are projects from undergraduates that look more promising and lively, filled with more passion than whatever "the next AI/YC/LinkedIn startup slop" I found nowadays. What is even more interesting was that all these projects either didn't use AI at all for coding, development, or minimally and only for debugging. It wasn't out of fear of skill atrophy, either. A few likes the challenge. Most of them wanted to control the code the way they wanted, and LLM or coding agents could not provide that to them. Three of the best projects I want to give a shout-out to are:
- Run for your life: a Pokemon Go-styled multiplayer running app
- Tasktix: a Trello-styled task manager. Free and open source. I appreciated the team for having a concrete programming standard. Because of the strict coding standard, coding agents can't really measure up so they hardly use them.
- Time to Love: a free to play game on Steam. Made by many developers, artists, musicians. No trace of AI in their development process, not in art, writing, or music.
With the newly graduating students booing at the school for daring to invite those "AI acolytes" and these innovative, passionate projects, I feel as if there is still hope and sanity left in this flaming pile of mess we call "our world".
