For better or worse, I’m pretty sure AI coding agents are not only capable of replacing developers in some distant future, they can do it now.
Below is a project that cost me about 10 USD in Deepseek API usage, about 10 days of sleep-deprived nights since I had to tell the agent what to do and test the outcomes (and, yes, to occasionally intervene and re-write things).
On that note, btw… Time estimates the modesl will, sometimes, give you, can be funny. It will tell you “it’s going to take a few weeks”, then it’ll be done in a few minutes.
So… Why not to use OpenClaw? I don’t trust it.
Why not to use Claude Code? Tbh, it’s simply way more expensive than DeepSeek (the model I used to essentially write the project below. Yes, it can basically code itself)
What about OpenAI? See “Claude” above.
Youll see a recording of how I’m asking the agent to describe the project, and it’s doing it on the fly. You can also see tool calling history etc. One thing it missed to mention, that, for the most part, this is running on Deepseek. It’s not hardcoded anywhere in the project, though, it’s just a model I’m using most of the times, so, from the project perspective, it’s run-time “configuration” rather than “architecture”.
Will I share this on github? That’s an interesting question… the intent is not to make this public – it’s my personal agent/assistant, it has remote runner capabilities so I can literally start the runner on one machine, open A-Gent in the browser from another, and it’ll be able to execute tools on the remote machine. This is a lot of potential exposure, and I don’t really want to deal with security issues or take responsibility for this thing by publishig it somewhere. It does what I need it to do for me, though.
And it’s not the point – the point is to illustrate what’s doable in just 10 nights, about $10 USD for Deepseek API calls, and quite a bit of testing/prompting/occasional hands-on development.
And if it were the only thing you could do… But it’s only an example. Those coding assistants are taking away the burden of having to “write code”. Yes they can make mistakes, yes you need to be accurate about what you are asking (to the point where the model can get confused simply because of the terminology), you probably do need some coding skills to jump in and connect the dots for the model every now and then when you can’t get it to do the right thing by adjusting the prompt. And yes, those tokens are getting more expensive as the projects grows, so you need to try to keep things “isolated” by design.
So will AI replace developers? Of course. The question is what we mean by that. AI will be spitting out smaller projects and changes all the time, just look at how Github has been having problems simply because of that sort of activity. However, AI does require guidance, and, as much as it’s replacing the “coding” part, I’d be nowhere close to having the same kind of personal assistant developed if I was not guiding AI through all the steps along the way, and that activity reuqired some weird combination of development/QA/BA/Architecture skills.