Coding Agents and Use Cases(justsitandgrin.im)
justsitandgrin.im
Coding Agents and Use Cases
https://justsitandgrin.im/posts/coding-agents-use-cases/
9 comments
I treat any article on the subject as corporate output to lay cover to say “we know AI” when in fact, they are at the same place their customers are. Where we all are.
I do like Antigravity a lot and I use it on a daily basis. Are you referring to that or something else? I do find their marketing weird though and seems like their internal teams are not yet fully aligned. (Jules, Antigravity, Gemini CLI, ...)
I've come across only a few of those. I'm not familiar with Antigravity. That's what I mean with poor marketing. I've tried out AI Studio, which is alright but not quite Codex level good as far as I can see.
Honestly this smells a bit like the situation they had a few years ago with a lot of different teams in Google inventing new chat and video call tools. All of which then flopped and got discontinued. Google exporting their own internal chaos. How many coding tools can you have in one company?!
OpenAI and Anthropic seem more focused currently. I mostly focus on OpenAI's codex because I don't want to maintain two subscriptions. Switching tools is a bit disruptive and not something I want to do every few days/weeks.
Honestly this smells a bit like the situation they had a few years ago with a lot of different teams in Google inventing new chat and video call tools. All of which then flopped and got discontinued. Google exporting their own internal chaos. How many coding tools can you have in one company?!
OpenAI and Anthropic seem more focused currently. I mostly focus on OpenAI's codex because I don't want to maintain two subscriptions. Switching tools is a bit disruptive and not something I want to do every few days/weeks.
Wouldn’t it make a lot of sense to have a single subscription that gives access to all coding agents, with the ability to switch between them based on preference or task? Constantly juggling tools and subscriptions is pretty disruptive.
20 euro per month is pretty hard to beat for Chat GPT Plus (which includes Codex). Right now a lot of this stuff in this space is highly experimental and constantly changing. I've been using codex web since before the summer, Codex cli for the last few months and Codex desktop app for the last few days. Before that, I was copy pasting blobs of code from Chat GPT and only looking at single code files.
The whole agentic coding revolution didn't really start moving until Claude Code rolled out almost a year ago now. Anthropic deserves kudos for that. Initially it was limited by small context sizes and wasn't that effective on large code bases. So, I stuck with ChatGPT and OpenAI; mainly because Claude for Desktop was a bit underwhelming and I felt OpenAI had their shit together a bit more in terms of UX, which I think matters at least as much as model quality for being able to effectively use AI. Arguably, since about the summer, Codex and Claude Code are well matched in terms of features/capabilities. Some prefer one or the other or use both. Coded had a reputation of being maybe slightly better with larger code bases. I don't think that is valid anymore as of the last few model releases.
Since about the gpt 5 generation of models released around last summer, I'm able to work on whole git repositories with codex. Our backend is about five years old; about 85K lines of code. Not huge but big enough that there was no way in hell LLMs were able to make sense of it before that. I only did the first big pull requests with codex on this backend in the last two months. This stuff is still very new.
I'm sure that with a lot of tool juggling and experimentation I could have gotten there a few weeks/months earlier But not much more. And I don't actually have time to be constantly fiddling with tools and trying out a lot of stuff. I don't need to be the first to try everything out.
The whole agentic coding revolution didn't really start moving until Claude Code rolled out almost a year ago now. Anthropic deserves kudos for that. Initially it was limited by small context sizes and wasn't that effective on large code bases. So, I stuck with ChatGPT and OpenAI; mainly because Claude for Desktop was a bit underwhelming and I felt OpenAI had their shit together a bit more in terms of UX, which I think matters at least as much as model quality for being able to effectively use AI. Arguably, since about the summer, Codex and Claude Code are well matched in terms of features/capabilities. Some prefer one or the other or use both. Coded had a reputation of being maybe slightly better with larger code bases. I don't think that is valid anymore as of the last few model releases.
Since about the gpt 5 generation of models released around last summer, I'm able to work on whole git repositories with codex. Our backend is about five years old; about 85K lines of code. Not huge but big enough that there was no way in hell LLMs were able to make sense of it before that. I only did the first big pull requests with codex on this backend in the last two months. This stuff is still very new.
I'm sure that with a lot of tool juggling and experimentation I could have gotten there a few weeks/months earlier But not much more. And I don't actually have time to be constantly fiddling with tools and trying out a lot of stuff. I don't need to be the first to try everything out.
The answer to “which coding agent should we standardise on?” is really simple: Don’t.
The tradeoffs of the different models are complicated and difficult to wrap your head around, and if you have the resources to try a bunch and form a conclusion, next week new models will come out and change the equation in small but difficult-to-understand ways again.
The solution is to ask your engineers which models they like, get them access to as many of those as you can, and expect their preferences to change and price that in.
“But I don’t have the budget to buy subscriptions to all the models my engineers want! And there are compliance issues with some of them!”
Note that I didn’t say you have to get access to all of them: as many as you can. And try to push the envelope as much as possible. Get creative. Perhaps give engineers a $200/mo AI coding budget and let them pick from a selection of subscriptions. You’re going to have different engineers using different AI coding tools, and if you refuse to let them, your competitors will.
Maybe in the future “standardizing on one coding agent” will be a thing that makes sense. But that time is not now.
The tradeoffs of the different models are complicated and difficult to wrap your head around, and if you have the resources to try a bunch and form a conclusion, next week new models will come out and change the equation in small but difficult-to-understand ways again.
The solution is to ask your engineers which models they like, get them access to as many of those as you can, and expect their preferences to change and price that in.
“But I don’t have the budget to buy subscriptions to all the models my engineers want! And there are compliance issues with some of them!”
Note that I didn’t say you have to get access to all of them: as many as you can. And try to push the envelope as much as possible. Get creative. Perhaps give engineers a $200/mo AI coding budget and let them pick from a selection of subscriptions. You’re going to have different engineers using different AI coding tools, and if you refuse to let them, your competitors will.
Maybe in the future “standardizing on one coding agent” will be a thing that makes sense. But that time is not now.
Are there security best practices for each of these tools or is that too early?
I don't have any use cases at all.
Business is definitely not booming.
Business is definitely not booming.
My views on this:
- the models don't matter as much any more. You can pick Gemini, Claude, or Codex and get similarish results. Especially if you are not a rocket scientist and aren't trying to boil the oceans and are just trying to complete relatively straightforward development tasks. You probably don't need the most expensive models on "high". You get a new one every few weeks.
- The UIs and plumbing around the models matter a lot more. To orchestrate agentic loops around these models is a lot of plumbing and work. Claude Code and Codex are the two "it just works" type options in this space. Claude had all the momentum earlier than OpenAI. But Codex is getting a lot of glowing reviews now. If you are more of a tinkerer, there are a bazillion tool and model combinations to try out that don't meaningfully move beyond what those two do out of the box. Your mileage may vary. I'm mostly a codex guy. Because I'm lazy and the OpenAI Plus subscription is a very good deal.
- Google is good at models but not that great at tools. Yet. And they are worse at marketing them. If find their stuff very confusing.
- Any advice on what is "best" here has a half life measured in weeks. It can all be different in a few weeks.