AI Coding Agents Are Already Commoditized (seangoedecke.com) 31
Software engineer Sean Goedecke argues that AI coding agents have already been commoditized because they require no special technical advantages, just better base models. He writes: All of a sudden, it's the year of AI coding agents. Claude released Claude Code, OpenAI released their Codex agent, GitHub released its own autonomous coding agent, and so on. I've done my fair share of writing about whether AI coding agents will replace developers, and in the meantime how best to use them in your work. Instead, I want to make what I think is now a pretty firm observation: AI coding agents have no secret sauce.
[...] The reason everyone's doing agents now is the same reason everyone's doing reinforcement learning now -- from one day to the next, the models got good enough. Claude Sonnet 3.7 is the clear frontrunner here. It's not the smartest model (in my opinion), but it is the most agentic: it can stick with a task and make good decisions over time better than other models with more raw brainpower. But other AI labs have more agentic models now as well. There is no moat.
There's also no moat to the actual agent code. It turns out that "put the model in a loop with a 'read file' and 'write file' tool" is good enough to do basically anything you want. I don't know for sure that the closed-source options operate like this, but it's an educated guess. In other words, the agent hackers in 2023 were correct, and the only reason they couldn't build Claude Code then was that they were too early to get to use the really good models.
[...] The reason everyone's doing agents now is the same reason everyone's doing reinforcement learning now -- from one day to the next, the models got good enough. Claude Sonnet 3.7 is the clear frontrunner here. It's not the smartest model (in my opinion), but it is the most agentic: it can stick with a task and make good decisions over time better than other models with more raw brainpower. But other AI labs have more agentic models now as well. There is no moat.
There's also no moat to the actual agent code. It turns out that "put the model in a loop with a 'read file' and 'write file' tool" is good enough to do basically anything you want. I don't know for sure that the closed-source options operate like this, but it's an educated guess. In other words, the agent hackers in 2023 were correct, and the only reason they couldn't build Claude Code then was that they were too early to get to use the really good models.
Re: (Score:2)
GitHub already has plenty of "terrible code"...written by humans.
If this makes any sense to someone (Score:2)
Re: (Score:1)
From what I've read, it means that the agents are interchangeable.
I've been using Cline.bot's VS Code extension with various agents and have found that Claude Sonnet has been most reliable. It is getting somewhat expensive but works reasonably well.
Re: (Score:2)
You can use any AI coding tool. It will work as a hyper-autocompletion agent. It will do all the basic stuff which is often useful. None of them have solved the problem that sometimes they will put out awful code and you will have to think. They are all basically the same.
That means that there's no real benefit from going for a proprietary one and you should target the one which is most clear and open about how it works and where it gets its training data from.
"Agentic" means "work as a software component t
Re: (Score:2)
Reminds me of sed(1).
Although I don't recommend the special case of reading one file on one side and outputting the same file on the other:)
Re:If this makes any sense to someone (Score:5, Insightful)
It means "Must keep AI hype going! Must pretent it is the only true thing! Must make more money!".
It can safely be ignored as total bullshit of the marketing variant.
Fuck "good enough" (Score:1)
If you write "good enough" using agents, don't expect me to fix it for you when it goes horribly wrong.
Re: (Score:3)
The literal meaning of "good enough" is "enough to get paid and forget it".
No bearing at all on fitness for a purpose or need of fixing.
Re: (Score:2)
Thanks for saying that. "Good enough" is a management-level bullshit that allows them to stretch "in progress" into "done".
Re: Fuck "good enough" (Score:1)
Re: (Score:3)
Hahah, yes. And remember that "good enough" looks massively different when you take security, usability and maintainability into account. I guess the next few years will get interesting for some enterprises that depend on software and quite a few will drown in a mountain of AI generated technological debt and die. Stupid people doing stupid things.
Re: (Score:2)
Perhaps then I'll come back to them and say "Sure, I can fix it, but to work with this mess my rate is 2k a day".
Re: (Score:3)
The point they all miss is that writing code which works was never the problem. Any junior dev can do it.
Software engineering always was about balancing tradeoffs, figuring integration points, ensuring long term maintainability, structuring for release and deployment, aligning design with roadmap, communication and collaboration, etc.
Maybe an AI can eventually get there, but your prompt will be way bigger than the code. I'd rather write the code.
For the rest, we already had cookiecutters and snippets.
Re: (Score:2)
You're absolutely right. In order to get anywhere near functional with AI you need to be so specific with the description of what you want AI to produce, that it undermines the whole undertaking. If you allow for any degree of ambiguity, AI will do all it can to satisfy your reqs at the cost of anything else that you've left unspoken - things that human beings take for granted but computers are unable to comprehend its importance.
I still get terrible results from "coding" agents (Score:5, Interesting)
I can only image what terrible code must be the norm in the places where the "coding agents" available today are considered "good enough".
Re: (Score:3)
You are not the only one. I am beginning to think that _everybody_ reporting great sucesses in this space is lying, deep in delusion or only doing very simplistic code (and struggling to do even that by themselves).
Re: I still get terrible results from "coding" age (Score:2)
The thing that AI lets them get around is bad management. They suddenly don't have to have aeeting about every little aspect of every little feature. They can just get AI to generate them some slop. For these types of management traps, it's going to lead to seriously accelerated production. But it's also going to lead to a lot of production failures, and then a return to slow processes.
Re: (Score:2)
On one hand, I agree with you that the quality is very poor. On the other, I do find AI tools (specifically, GitHub Copilot) useful and time-saving. I almost always have to fix what it generates, but it still saves me time.
Some specific areas of success include:
- SQL commands that manipulate XML or JSON database blobs
- XPATH or JSONPATH generation
- Converting jQuery ajax to async/await with fetch
- Generating unit test skeletons
In each of these cases, you do have to know what you're doing, and you have to be
Hahahaha, no (Score:2)
This is just more assholes portraying things they profit from as "inevitable", trying to keep the hype going and delaying the point where enough people realize that LLMs are not the revolution they are advertized as. You know, like in _all_ other AI hypes before.
And "just" better base models? Good luck with that. It is like saying space travel "just" needs a cheap and reliable FTL engine and we are all set. True, but meaningless.
On A Long Enough Timeline (Score:2)
This all fails. Everyone gets replaced with AI, no one has a job, no one has money to support companies.
It's just they get to screw us first.
We need an AI ban. This is not going to be a good thing for society. It already isn't. People are going to die because of bullshit decisions made by AI...likely already have. When the black box is making the decision and no one lets you look in the black box....is there really a black box?
We're all fucked. Congrats. It's only going to get worse before it gets better.
Re: (Score:2)
Your prediction is like predictions from the 20th century, that calculators would destroy the field of mathematics, or that chess computers would destroy the game of chess. Or more recently, that Google Maps would destroy people's ability to navigate, or that Waymo would obliterate Uber and taxi services. All these technologies impacted the various fields in significant ways, but did not destroy them.
As a daily user of AI, I know that we are a LONG way from being displaced by AI. And even to the extent that
I'll be honest (Score:2)
AI marketplace (Score:2)
The IDE I use offers 25 backend AI agents to pick from in a dropdown menu. All of them are either free or very cheap except for Claude Sonnet 4.0, which is reportedly the best but it burns through credits. All of them are probably operating at a loss.
Re: (Score:2)
The IDE I use offers 25 backend AI agents to pick from in a dropdown menu. All of them are either free or very cheap except for Claude Sonnet 4.0, which is reportedly the best but it burns through credits. All of them are probably operating at a loss.
The future will be locally run models.
Only problem is that to promote cloud models, the hardware is being slow walked to do local models.
There is just so much consolidation and conflicts that AI development is being done in a very investment friendly way rather than the most efficient way.
So it won't last (Score:2)
AI is a technology that is going to very quickly belong to a handful of super wealthy companies. The AI's dependency on training data guarantees market consolidat