Agentic coding: does it really work, or is it just a skill issue on my side?

[eamarce23]

So I’ve been working lately with different coding agents, and I’ve noticed something interesting.

As projects grow, even when I improve the context, guardrails, specs, hooks, and context-management techniques, I still sometimes see glitches in the code:

* Files that were not supposed to be touched get modified.
* Previously working features get deleted.
* New issues get introduced.

To be clear, these don’t reach production. We still review changes, run tests, and promote carefully through higher environments.

But we do spend time finding and fixing these issues, whether by re-prompting the agent, making the changes ourselves, or using a mix of both.

So I’m curious:

Has anyone else experienced this with coding agents?

Or, as people often say on social media, is this just a skill issue on my side? 😂
I’m still learning, so I’m genuinely open to that possibility.

Being used to a bit more precision at work, this has left me a bit confused and, honestly, a little anxious haha.

Would love to hear how others are handling this.

Thanks for sharing your experience 🙂

[Vitaliy-R]

Would you also share which coding agents you use and which programming languages are used in these projects? Thanks!

[eamarce23]

Sure thing, Vitaly.

We have one biggie where we are working with Python for backend and long running processes and Javascript (Next + Node.js). We have used Claude Code and Cursor for that one in particular. 

We have smaller apps with SAP CAP + Fiori freestyle. Here we used Claude Code mainly. I had the opportunity to try Joule for Developers in the past 🙂 but not directly working with it in our pipeline at the moment.

Thanks for replying 🙂 

[DWS123]

Hi eamarce23  - its not you, it is the coding agents.  I have been dealing with this for some time now.  I do not know specifically what agents you are using, in which environment, or for what code.  I have found that different models work differently and have different strengths and weakness.  I will not say better or worse but different.  Having said that, for the very reason you mentioned, I have officially "fired" one model entirely.  It was forgetting and just getting things wrong.   For the most part, I have found that the issues are "memory" issues.   I would suggest that you ask the agent what you can do to improve the work.    I have done this with several agents.   One of the agents suggested that a clear design be agreed to up front and is managed in a methodical way.   I guess there is no getting away from that.  It also suggested periodic design versioning, that becomes an "approved" version and the model cannot go back and change it, without permission.    One model got so back that I had to institute a rule that I needed to review every change, but even then it still made mistakes.   Another model, when I asked how to stop these types of issues, suggested that I periodically create "state documents" and then take that document into a new chat and start fresh.  According to the model it has trouble the longer the chat gets.   This seems to help.    Some agents have come back and just said "sorry that's on me" and has moved on.  Does not mean much coming from something that cant actually feel regret.  But at least you know its not you.    I hope this helps.   

[eamarce23]

Thank you for your input, @DWS123 🙂 Yes, I saw floating around the recommendation to ask the agent directly, so I will try that. One thing I try to do is to start a new session for every feature I need to work on. I was reading a nice book that explains how, in the end, the skills, and other parts of the context compound to a bigger context for the agent to work. I'm also trying to take that into account. So far, I've encountered the technology to be a good booster for different tasks, but still being far from fully autonomous coding. I will continue looking at more results and I will come back and share as well. 

Wondering if anybody else is experiencing the same as we are, and what other things might be helpful to consider 😀

[Private_Member_49934]

Which book are you reading?

[eamarce23]

🙂I had been looking for a resource like this for a while, had mainly done Youtube videos and documentation in the past: 

Agentic Coding with Claude Code - Eden Marco - Packt
https://www.packtpub.com/en-us/product/agentic-coding-with-claude-code-9781806022588

So far liking it, mainly because you get to see what is behind different pieces in Claude Code.