From ChatGPT to Coding Agents: Best Friend or Worst Enemy?

I wrote the first version of this article in 2023, when "AI coding" meant pasting snippets into ChatGPT and hoping. The question was whether that chatbot would help developers or hurt them.

The question aged in a funny way. Nobody pastes snippets anymore. AI now lives inside the editor and the terminal, works as agents that read your whole codebase, run your tests, and open pull requests. I build software for a living and most days AI writes a large share of my code.

So here is the updated answer, from someone who ships production systems this way, including AI products themselves.

The best friend case is real

Speed is not hype. Features that took me a week in 2022 take a day or two now. Not because the AI is magic, but because it removes the slow parts: boilerplate, wiring, test scaffolding, reading unfamiliar library docs, the hundredth CRUD endpoint.

Breadth changed what one person can ship. I work across mobile apps, backends, databases, infrastructure, and the AI layer on top. Keeping all of that in one head used to be the hard limit of solo work. With AI handling recall and syntax, one experienced developer can now cover a surface that used to need a small team.

It raises the floor on quality, if you let it. More tests get written because writing them is cheap. Edge cases get handled because asking "what did I miss" costs nothing. Refactors happen because they are no longer a whole afternoon.

The worst enemy case is also real

I have seen the failure modes up close, in my own work and in codebases I get called into.

AI code looks right more convincingly than wrong code ever did. It compiles, reads clean, and handles the happy path. The bug is in the edge case the model never considered, wearing the confidence of well-formatted code. Plausible-but-wrong is the default failure mode, not the exception.

Judgment atrophies quietly. If you accept whatever the agent produces, you stop noticing bad architecture as it accumulates. Six months later the codebase works but nobody can explain why, and every change breaks something. I have been handed exactly these projects to rescue. The "vibe coding" phase, building by accepting AI output without understanding it, produces prototypes that demo well and collapse in production.

Security and data handling need an adult in the room. Models happily write string-concatenated SQL, log secrets, and skip authorization checks if the prompt does not force the issue. The model optimizes for "works in the demo", not "safe with real users".

The rules that keep it a best friend

None of this made me use AI less. It made me use it with discipline. These are the actual rules I work by:

1. Plan first, generate second. The approach gets agreed before code gets written. AI executing a clear plan is excellent. AI inventing architecture on the fly is how you get a codebase nobody owns.

2. Everything runs locally before it ships. Generated code is a draft until it has been executed and verified. No exceptions for code that "obviously works".

3. Every bug becomes a test. AI makes writing regression tests nearly free, so there is no excuse. Fix it once, pin it down forever.

4. Read what you merge. Not every line with equal attention, but every decision. If I cannot explain why the code is shaped the way it is, it does not go in.

5. The human owns the failure. When something breaks in production, "the AI wrote it" is not a sentence I ever get to say to a client. Pricing that responsibility in changes how carefully you review.

If those rules sound like ordinary engineering discipline, that is the point. AI did not change what good software practice is. It changed the cost of ignoring it, because you can now produce bad code at unprecedented speed.

Best friend or worst enemy?

Both, and the deciding factor is not the tool. Paired with someone who keeps their engineering judgment engaged, AI coding tools are the biggest productivity gain of my career. Handed to someone who outsources their thinking, the same tools generate technical debt faster than any human could type it.

The developers who struggle are not the ones who refused AI. They are the ones who let it do the judging. Keep the judgment, delegate the typing, and it is the best colleague you have ever had.