Gen AI: Tutor, Helper or Cheater? The Conductor decides.

A year back, I decided to start a passion project: building an e-Invoice application. I picked up a new language, Python. On a typical self-learning path, I started going through a bunch of tutorials, whether in written articles, ebooks, or YouTube videos. From the very basics of data types, loops, dictionaries, and sets, through intermediate topics like file and exception handling and object-oriented programming, all the way to advanced topics such as decorators and concurrency. It took me a few weeks to get my bearings before I started developing my project.

There was a lot of joy in those weeks of learning and building. Whether it was figuring out how to form a JSON request from dictionaries, or solving that darn UI “freeze” issue when doing I/O with threading. I’m sure this resonates with those of you who do development. You know that euphoric “Aha!” moment when you finally crack it, sometimes with your fist thrust into the air. Of course, the path from problem to solution was never quick. Simple problems might take a couple of hours; more complex ones easily took days — and that’s assuming you had a clear, focused mind, not one cluttered by the other 101 bugs and new features on your TODO list.

Then came Generative AI (Gen AI). AI tools, harnesses, and frameworks were being made available to the public at breakneck speed. Every other week, it seemed, there were new GA (General Availability) releases for large language models (LLMs — ChatGPT, Gemini, Gemma, DeepSeek), AI harnesses (OpenClaw, Claude Code, etc.), and frameworks (LangGraph, Microsoft Agent Framework, etc.). It was very hard to ignore all of this. I started using Gen AI in my project to explore its potential.

Wow, what a thrill ride! I used OpenClaw and integrated it with DeepSeek and Google Gemini as the LLMs, predominantly for code generation in my project. The strength of OpenClaw lies in its agents. You can create specialized agents that focus on specific tasks. For example, you can have a coding agent that handles all the implementation, or a tester agent that runs functional and regression tests. In practice, you can assemble an entire development team — architect, developer, tester, and documentation-writer agents. As you can imagine, I had many “conversations” with my agents, strategizing implementation approaches on topics like building efficient dictionary-to-JSON and XML classes, or streamlining helper classes. They taught me domains I wasn’t familiar with, like certificate signing and encryption in general. When I found a bug, all I needed to do was hand over enough log dumps, and the agent would walk me through the root cause and possible solutions, step by step. It might take a few attempts; nevertheless, after a few rounds of back-and-forth, we’d fix it.

Now, once you get used to it, Gen AI can be very addictive. A feature that normally takes a couple of days to complete, Gen AI can finish in a matter of hours. Bugs that surface late in the evening (for some reason, they always do), which used to keep you up till the wee hours of the morning, can now be fixed quickly with Gen AI as your compatriot. You just need to provide the right logs and lead the troubleshooting effectively. Over time, your reliance on Gen AI grows without you realizing it. At some point, you’ve slipped into “vibe-coding” mode without even noticing.

There’s been plenty of debate on how useful Gen AI really is. It has proven very good at its tasks, and it will only get better in the coming years, if not months. The momentum is so strong that many have concluded it will take away jobs (and to a certain extent, it already has), creating fear across the workforce. It certainly doesn’t help that big players — Google, Meta, Microsoft, Oracle, to name a few, have all announced layoffs and reportedly redirecting the savings toward AI-related costs (compute, data centres, LLMs). Add to that the sensational, click-bait headlines on social media about how Gen AI is killing jobs, and things look gloomy for job seekers, especially fresh graduates and junior workers.

Indeed, Gen AI can seem like the perfect architect, developer, tester, or documentation writer (however you set up your agents), but it still makes mistakes. We must remember that an LLM (the “brain” the agents use) is only as good as its training data. LLMs learn through pattern recognition, and by no means do they have the same cognition we humans have. That’s exactly why, whenever you use ChatGPT, Gemini, etc., you’re greeted with a friendly reminder along the lines of: Generative AI can make mistakes or overlook recent updates. Do not rely solely on this response; please check with the relevant party, authority, regulator and so on. Sound familiar? Gen AI will hallucinate (translation: it will make up a story ) when it’s unsure. It will be utterly convincing, assuring you the facts it gives are correct, when in fact the opposite is true. I’ve found myself “arguing” with my agents over development approaches I know for a fact won’t work. It takes a few rounds of back-and-forth to get them to “understand” why it will fail. It reminds me of dealing with that one stubborn engineer who thinks he’s never wrong.

I see Gen AI as a very good tool, part of my workflow and a genuine productivity boost. But it isn’t foolproof. The agents I build are only as good as how well I define them. Yeah, someone has to set up the agent (i.e., configure the AGENTS.md). There are frameworks that provide out-of-the-box agents nowadays; still, to make them work well for you, fine-tuning is required. It also means you need to hold the big picture and direct the agents to handle the nitty-gritty details. Think of it like an orchestra. You are the conductor; your agents are the ensemble. You lead the symphony, from the first movement to the fourth. The moment you go on autopilot (a.k.a. full vibe-coding), getting things done without understanding what the agents actually did, you're setting yourself up for disaster. No clue why a feature isn’t working, or how a bug got there in the first place.

How you use Gen AI is your call... tutor, helper or cheater? And, as with any tool a team adopts, the question becomes: how do we measure its effectiveness and productivity? That's a topic for another time.