HN Reader

I'm one of the creators of OpenHands (fka OpenDevin). I agree with most of what's been said here, wrt to software agents in general.

We are not even close to the point where AI can "replace" a software engineer. Their code still needs to be reviewed and tested, at least as much as you'd scrutinize the code of a brand new engineer just out of boot camp. I've talked to companies who went all in on AI engineers, only to realize two months later that their codebase was rotting because no one was reviewing the changes.

But once you develop some intuition for how to use them, software agents can be a _massive_ boost to productivity. ~20% of the commits to the OpenHands codebase are now authored or co-authored by OpenHands itself. I especially love asking it to do simple, tedious things like fixing merge conflicts or failing linters. It's great at getting an existing PR over the line.

It's also important to keep in mind that these agents are literally improving on a _weekly_ basis. A few weeks ago we were at the top of the SWE-bench leaderboard; now there are half a dozen agents that have pulled ahead of us. And we're one launch away from leapfrogging back to the top. Exciting times!

https://github.com/All-Hands-AI/OpenHands

Those “how I feel about Devin after using it” comments at the bottom are damning, when you compare them to the user testimonials of people using cursor.

Seems to me that agents just aren’t the answer people want them to be, just a hype wave obscuring real progress in other areas (eg. MCST) because they’re easy to implement.

…but really, if things are easy to implement, at this point, you have to ask why they haven’t been done yet.

Probably, it seems, because it’s harder to implement in a way that’s useful than it superficially appears…

Ie. If the smart folk working on Devin can only do something of this level, anyone working on agentic systems should be worried, because it’s unlikely you can do better, without better underlying models.

I think one of the big problems with Devin (and AI agents in general) is that they're only ever as good as they are. Sometimes their intelligence feels magical and they accomplish things within minutes that even mid level or senior software engineers would take a few hours to do. Other times, they make simple mistakes and no matter how much help you give, they run around in circles.

A big quality that I value in junior engineers is coachability. If an AI agent can't be coached (and it doesn't look like it right now), then there's no way I'll ever enjoy using one.

One thing that surprised me a little is that there doesn't seem to be an "ask for help" escape hatch in it - it would work away for literally days on a task where any human would admit they were stuck?

One of the more important features of agents is supposedly that they can stop and ask for human input when necessary. It seems it does do this for "hard stops" - like when it needed a human to setup API keys in their cloud console - but for "soft stops" it wouldn't.

By contrast, a human dev would probably throw in the towel after a couple of hours and ask a senior dev for guidance. The chat interface definitely supports that with this system but apparently the agent will churn away in a sort of "infinite thinking loop". (This matches my limited experience with other agentic systems too.)

I remain sceptical about the "Planet Tracker"-task. The task was to debunk claims about historical positions of Jupiter and Saturn. If the task was to find those planets were NOT in a certain (claimed) position an erroneous program would still appear to "debunk" the claims. Did they check if Devin's code's calculated positions were actually correct? Did they check in some NASA-database? If Devin gave arbitrary positions for the planets it's much more likely that they're different than any claim and appear to debunk it.

What model does Devin use? How would it change if it used o1 or even o3 for times when it gets stuck?

IE. Generate the initial code using GPT4o/Claude 3.5, then start testing the code, when it gets stuck, use o1/o3 to help.

This feels a bit like AI image generation in 2022. The fact that it works at all is pretty mindblowing, and sometimes it produces something really good, but most of the time there are obvious mistakes, errors, etc. Of course, it only took a couple more years to get photorealistic image outputs.

A lot of commenters here seem very quick to write off Devin / similar ideas permanently. But I'd guess in a few years the progress will be remarkable.

One stubborn problem – when I prompt Midjourney, what I get back is often very high-quality, but somehow different than what I expected. In other words, I wouldn't have been able to describe what I wanted, but once I see the output I know it's not quite right. I suspect tools like this will run into similar issues. Maybe there will be features that can help users 'iterate' quickly.

I also wrote my first impressions on Devin, more focused on the user experience and analysis of its capabilities (with lots of screenshots):

https://thegroundtruth.substack.com/p/devin-first-impression...

The whole idea of Devin is pointless and doomed to fail in my humble opinion, big tech will be quite capable on delivering A.I agents / assistants - very soon. I don't think wrappers over other people's LLMs like Devin make a lot of sense. Can someone help me understand what's the value proposition / moat of this company?

> Even more concerning was Devin’s tendency to press forward with tasks that weren’t actually possible. (...)

> Devin spent over a day attempting various approaches and hallucinating features that didn’t exist.

One of the big problems of GenAI is its inability to know what they don't know.

Because of that, they don't ask clarifying questions.

Humans, in the same situation, would spend a lot of time learning before they could be truly productive.

Sounds exactly like my experience with the “agents” about a year ago. Autogpt or whatever it was called. Works great 1% of the time and the rest it gets stuck in the wrong places completely unable to back out.

I’m now using o1 or Claude Sonnet 3.5 and usually one of them gets it right.

An engineer that thinks it knows everything (but doesn't) and can't self-correct is about the worst combo I can think of.

Disclosure: Working on a company in the space and have recently been compared to Devin in at least one public talk.

Devin has tried to do too much. There is value in producing a solid code artifact that can be handed off for review to other developers in limited capacities like P2s and minor bugs which pile up in business backlogs.

Focusing on specific elements of the development loop such as fix bugs, add small feature, run tests, produce pull request is enough.

Businesses like Factory AI or my own are taking that approach and we're seeing real interest in our products.

The thing with AI agents I tend to find is they reveal how much heavy lifting the dev is actually doing.

A personal example, my best use out of AI so far has been cases where documentation was poor to nonexistent, and Claude was able to give me a solution. But the thing is, it wasn't a working solution, nowhere close, but it was enough for me to extrapolate and do my own research based on the structure, classes and functions it used. Basically, it gave me somewhere to start from. Whether that's worth the social, economic and environmental problems is another story.

"Even more telling was that we couldn’t discern any pattern to predict which tasks would work."

I think this cuts to the core of the problem for having a human in the loop. If we cannot learn how to best use the tool from repeated use and discern some kind of patterns of best and worst practices then it isn't really a tool.

No matter what happens with Devin specifically, I think this is a really important topic and I enjoy reading updates on this kind of review every time.

Please keep them coming.

Also trialed Devin, it's quite impressive when it understands the code formatting and local test setup, producing well formatted and test case passing code, but it seems to always add extraneous changes beyond the task that can break other things. And it can't seem to undo those changes if you ask. So everything requires more cleanup. Devin opened my eyes to the power of agentic workflows with closed loop feedback, and the coolness of a slack interface, but I am gonna recommend cancelling it because it's not actually saving time and it's quite expensive.

I've been experimenting with code gen on and off for the last 18 months, and find this exactly in line with my experience.

As someone who uses AI coding tools daily and has done a fair amount of experimentation with different approaches (though not Devin), I feel like this tracks pretty well. The problem is that Devin and other "agentic" approaches take on more than they can handle. The best AI coders are positioned as tools for developers, rather than replacements for them.

Github Copilot is "a better tab complete". Sure, it's a neat demo that it can produce a fast inverse square root, but the real utility is that it completes repetitive code. It's like having a dynamic snippet library always available that I never have to configure.

Aider is the next step up the abstraction ladder. It can edit in more locations than just the current cursor position, so it can perform some more high-level edit operations. And although it also uses a smarter model than Copilot, it still isn't very "smart" at the end of the day, and will hallucinate functions and make pointless changes when you give it a problem to solve.

If you're going to compare other tooling, I'm curious to know what you think of our long term goals: https://github.com/charperbonaroo/robtherobot/issues/2

So for anyone who doubted SWE-BENCH's relevance's to typical tasks, it seems that its stated 13.86% almost exactly matches this 3 successes out of 20 pilot outcome.

We're not quite there yet, but all of these issues seem to be solvable with current tech by applying additional training and/or "agentic" direction. I would now expect pretty much the textbook disruptive innovation process over the next decade or so, until the typical human dev role is pushed to something more akin to the responsibilities of current day architects and product managers. QA engineering though will likely see a big resurgence.

I’ve used Cursor a lot and the conclusion doesn’t surprise me. I feel like I’m the one *forcing* the system in a certain direction and sometimes an LLM gives a small snippet of useful code. Sometimes it goes in the wrong direction and I have to abort the suggestion and force it into another direction. For me, the main benefit is having a typing assistant which can save me from typing one line here and there. Especially refactorings is where Cursor shines. Things like moving argument order around or adding/removing a parameter at function callsites is great. Saved me a ton of typing and time already. I’m way more comfortable just quickly doing a refactoring when I see one.

AI in general doesn’t understand scope

The assumption with low-code tooling was that AI is so good at writing actual code in a way that it will make low-code tools redundant. Spending time with Windsurf, Cursor, and a bunch of VSCode extensions, while it was so impressive to see new projects being created autonomously, asking for new requirements or fixing bugs after >10 iterations was more complex.

I had to audit the code and give specific directions on how to restructure the code to avoid getting stuck when the project gets more complex. That makes me think autonomous agents will do much better on low-code tools, as their restrictions ensure the agent is on track. The problem with low-code tools is that they also get more complicated to scale after maybe like >200 iterations. (for a medium-sized project, on average 6 months)

This only reinforces my bias against AI agents. At this point, they are mostly just hype. I believe that for AI to replace a junior, we would need to achieve at least near-AGI, and we are far from that.

I'm sure a lot of folks in these comments predicted these sorts of results with surprising accuracy.

Stuff like this is why I scoff when I hear about CEOs freezing engineering hiring or saying they just don't need mid-level engineers anymore because they have AI.

I'll start believing that when I see it happening, and see actual engineers saying that AI can replace a human.

I am long AI, but I think the winning formula is small, repetitive tasks with a little too much variation to make it worth it (or possible) to automate procedurally. Pulling data from Notion into Google sheets, like these folks did initially, is probably fine. Having it manage your infrastructure and app deployments, likely not.

I'm working on AI assistant in Python notebook. It aims to help with data science tasks. I'm not using it to do a full analysis. It will fail. What I ask is to create a code snippet for my next step in the analysis. Many times I need to manually change the code, but it is fine because LLM speed-up my coding a lot. And it is really fantastic in writing matplotlib code for visualization. I don't remember all matplotlib syntax to change axis labels, add annotations or change style, and LLM really can handle it good, in impressive speed.

Most the problems you mentioned will likely be solved with the next iterations of Devin or similar product.

I can say that because I work daily with Claude as an agent over mcp, and the problems you mentioned feel very familiar.

Based on the type of the issues you mentioned, Devin isn't likely using o1 yet. A workflow like o1 for planning, Claude for Coding, o1 for review, etc., would work better.

The problems you mentioned: ssh-key issue unrelated to script, code not following existing patterns or themes, instructions not being followed, extra abstractions, etc., fall into that category.

Some of the issues are likely due to context length problem. For example, LLM doesn't work well with jupyter notebook because of extra junk in ipynb, which will likely remain a problem.

I can’t believe they named it after Devin Franco - guess it can take a lot of load!

I saw few people around testing and it is quite disappointing. Sometimes a task might take forever and deliver a bad result or fail completely.

It seems it is targeting few specific problems and whatever else is just too hard. I also think that, thought it is expensive, it is cheap for the technology behind it and it won't be able to keep that price for long

Now is the time for us to hold seemingly contradictory propositions: A child born today will live to see 99% of all computer code written by artificial intelligence, but the current AI boom is massively overcapitalized.

At some point people are going to realize that using these LLM AIs is a communications problem, and by that I mean the reason various attempts to use them fail is because they are not being effectively told what to do, vague and implied requests are not enough for a inhuman statistical construct to grasp what you're asking without clearer more details and more specific instructions.

Honestly, i have been bitten so many times by LLM hallucinations when I work in parallel with the LLM, I wouldn't trust it autonomously running anything at all. If you have tried to use imaginary APIs, imaginary configuration and imaginary cli arguments, you know what I mean

Do you have good references about using AI coding assistants?

Techniques of prompt engineering help a lot, but I really think there will be created a body of knowledge about how to use, what's the good contexts of use, and good heuristics. They are a valuable tool, but I feel it's possible to extract more value.