My experience is also that they're very different from writing code myself.
There are ways to use the tool better and worse, the worst is to blindly fire off the agent and commit the changes without reading them: This *only* works for very small projects, though it often *does* work for those projects.
For bigger projects, you need to use (and enforce use of) the right architecture, and read pull requests rather than saying "LGTM" and clicking accept. Pick the right way to split the project up into discrete components with clear boundaries between them, make sure that the LLM's commits never introduce surprise coupling outside those boundaries, and make sure the unit tests aren't ridiculously brittle, they can be OK.
"Make sure to read the commits" seems to be enough for your points 2/3, at least for me. But everyone thinks a bit differently, YMMV. It also means I almost immediately catch most of the (20%? 10%?) of the time the LLM does something wrong, though not all of them.
I find I still have deep focus. If anything, the problem for me is they're so addictive I am tempted to keep working and working rather than rest, a problem I last faced over 20 years ago when I was working on a personal project that ended up not going anywhere.
I do read the commits and really try to understand what the model is trying to do. However, I think this conflicts against what we are really trying to do here:
- We want to code faster so we use LLMs
- We want to still be in charge, so we spend a lot of time reading and understanding every single commit and review PRs.
Before, we used to build a mental model of the code with every new line of code we wrote. That took time, but the result was a very solid understanding. Now we try to build the mental model by just looking at what the LLM produces. Similar to how students who spend most of their time practicing what they are studying with actual exercises achieve higher grades than those who just read the book, we are now at a disadvantage against our old selves who made the effort of writing the code.
Moreover, an LLM produces code a lot faster than we can read and understand. If we want to achieve an actual speedup in code output, there's no way we can read and understand everything and be faster than before. We have to sacrifice the understanding of some parts. This compounds over time and we quickly get into a place of total reliance on LLMs.
Regarding addiction: I also think these devices are addictive. They do open a lot of new doors for us builders. But I think hour grey matter hardware is not built of this speed. Give it time, the more project ideas you start spinning up, the sooner you'll reach burnout.
> Similar to how students who spend most of their time practicing what they are studying with actual exercises achieve higher grades
This is spot on.
> build a mental model of the code with every new line of code
This idea has been explored in depth in the Peter Naur's excellent paper, "Programming as theory building" [1] (also extensively discussed on HN [2]). If you haven't read it yet, I bet you will enjoy it a lot. :)
I find it painful as well. I actually find agentic coding counterproductive. I typically use LLMs the “old way”: web chat interface, asking questions and searching. I write the code by hand. This makes me 2-3x more productive than before (I guess with agents I could be 5-10x more productive but the price to pay is not worth it)
I don't really think so. When writing code with AI, I feel like we're doing two different jobs at once: writing documentation and writing code. Even though they're technically two sides of the same coin, it gets really frustrating.
Besides, no matter how good our design docs are, AI just can't grasp the business context outside of what's written. Plus, AI loves to modify things using band-aid patches, making it incredibly hard to keep the code clean. The design logic often gets messy, and you ultimately lose true logical consistency.
My current workaround is to provide detailed architecture and design principles upfront, and also add comments at the top of each file to clearly define its scope. That seems to help a bit.
But my biggest headache right now is that AI struggles to adapt to the latest APIs—it actually likes to revert the new APIs I've just updated back to the older, deprecated ones.
Yes! Yes! Yes! The band-aid patches! I really hate that.
Regarding old API regression, I think that's because the LLM sees a lot more examples of deprecated APIs throughout the code, than new replacements you just introduced. You could try forcing it to immediately replace old calls to deprecated endpoints, then actually removing the definition completely, so that you reduce the number of bad examples and increase that of good examples
My experience is also that they're very different from writing code myself.
There are ways to use the tool better and worse, the worst is to blindly fire off the agent and commit the changes without reading them: This *only* works for very small projects, though it often *does* work for those projects.
For bigger projects, you need to use (and enforce use of) the right architecture, and read pull requests rather than saying "LGTM" and clicking accept. Pick the right way to split the project up into discrete components with clear boundaries between them, make sure that the LLM's commits never introduce surprise coupling outside those boundaries, and make sure the unit tests aren't ridiculously brittle, they can be OK.
https://www.youtube.com/watch?v=grPtnrOTP_4
"Make sure to read the commits" seems to be enough for your points 2/3, at least for me. But everyone thinks a bit differently, YMMV. It also means I almost immediately catch most of the (20%? 10%?) of the time the LLM does something wrong, though not all of them.
I find I still have deep focus. If anything, the problem for me is they're so addictive I am tempted to keep working and working rather than rest, a problem I last faced over 20 years ago when I was working on a personal project that ended up not going anywhere.
I do read the commits and really try to understand what the model is trying to do. However, I think this conflicts against what we are really trying to do here: - We want to code faster so we use LLMs - We want to still be in charge, so we spend a lot of time reading and understanding every single commit and review PRs.
Before, we used to build a mental model of the code with every new line of code we wrote. That took time, but the result was a very solid understanding. Now we try to build the mental model by just looking at what the LLM produces. Similar to how students who spend most of their time practicing what they are studying with actual exercises achieve higher grades than those who just read the book, we are now at a disadvantage against our old selves who made the effort of writing the code.
Moreover, an LLM produces code a lot faster than we can read and understand. If we want to achieve an actual speedup in code output, there's no way we can read and understand everything and be faster than before. We have to sacrifice the understanding of some parts. This compounds over time and we quickly get into a place of total reliance on LLMs.
Regarding addiction: I also think these devices are addictive. They do open a lot of new doors for us builders. But I think hour grey matter hardware is not built of this speed. Give it time, the more project ideas you start spinning up, the sooner you'll reach burnout.
> Similar to how students who spend most of their time practicing what they are studying with actual exercises achieve higher grades
This is spot on.
> build a mental model of the code with every new line of code
This idea has been explored in depth in the Peter Naur's excellent paper, "Programming as theory building" [1] (also extensively discussed on HN [2]). If you haven't read it yet, I bet you will enjoy it a lot. :)
[1] https://gwern.net/doc/cs/algorithm/1985-naur.pdf
[2] https://hn.algolia.com/?q=Programming+as+theory+building
I find it painful as well. I actually find agentic coding counterproductive. I typically use LLMs the “old way”: web chat interface, asking questions and searching. I write the code by hand. This makes me 2-3x more productive than before (I guess with agents I could be 5-10x more productive but the price to pay is not worth it)
I don't really think so. When writing code with AI, I feel like we're doing two different jobs at once: writing documentation and writing code. Even though they're technically two sides of the same coin, it gets really frustrating.
Besides, no matter how good our design docs are, AI just can't grasp the business context outside of what's written. Plus, AI loves to modify things using band-aid patches, making it incredibly hard to keep the code clean. The design logic often gets messy, and you ultimately lose true logical consistency.
My current workaround is to provide detailed architecture and design principles upfront, and also add comments at the top of each file to clearly define its scope. That seems to help a bit.
But my biggest headache right now is that AI struggles to adapt to the latest APIs—it actually likes to revert the new APIs I've just updated back to the older, deprecated ones.
Yes! Yes! Yes! The band-aid patches! I really hate that.
Regarding old API regression, I think that's because the LLM sees a lot more examples of deprecated APIs throughout the code, than new replacements you just introduced. You could try forcing it to immediately replace old calls to deprecated endpoints, then actually removing the definition completely, so that you reduce the number of bad examples and increase that of good examples
So don't use coding agents? No one is forcing you, are they?