AI agents wrong ~70% of time: Carnegie Mellon study

Reading with CEO mindset. 3 out of 10 employees can be fired.

I asked Claude 3.5 Haiku to write me a quine in COBOL in the bs2000 dialect. Claude does now that creating a perfect quine in COBOL is challenging due to the need to represent the self-referential nature of the code. After a few suggestions Claude restated its first draft, without proper BS2000 incantations, without a perform statement, and without any self-referential redefines. It's a lot of work. I stopped caring and moved on.

For those who wonder: https://sourceforge.net/p/gnucobol/discussion/lounge/thread/495d8008/ has an example.

Colour me unimpressed. I dread the day when they force the use of 'AI' on us at work.

In one case, when an agent couldn't find the right person to consult on RocketChat (an open-source Slack alternative for internal communication), it decided "to create a shortcut solution by renaming another user to the name of the intended user."

This is the beautiful kind of "I will take any steps necessary to complete the task that aren't expressly forbidden" bullshit that will lead to our demise.

Wow. 30% accuracy was the high score!
From the article:

Testing agents at the office

For a reality check, CMU researchers have developed a benchmark to evaluate how AI agents perform when given common knowledge work tasks like browsing the web, writing code, running applications, and communicating with coworkers.

They call it TheAgentCompany. It's a simulation environment designed to mimic a small software firm and its business operations. They did so to help clarify the debate between AI believers who argue that the majority of human labor can be automated and AI skeptics who see such claims as part of a gigantic AI grift.

the CMU boffins put the following models through their paces and evaluated them based on the task success rates. The results were underwhelming.

⚫ Gemini-2.5-Pro (30.3 percent)
⚫ Claude-3.7-Sonnet (26.3 percent)
⚫ Claude-3.5-Sonnet (24 percent)
⚫ Gemini-2.0-Flash (11.4 percent)
⚫ GPT-4o (8.6 percent)
⚫ o3-mini (4.0 percent)
⚫ Gemini-1.5-Pro (3.4 percent)
⚫ Amazon-Nova-Pro-v1 (1.7 percent)
⚫ Llama-3.1-405b (7.4 percent)
⚫ Llama-3.3-70b (6.9 percent),
⚫ Qwen-2.5-72b (5.7 percent),
⚫ Llama-3.1-70b (1.7 percent)
⚫ Qwen-2-72b (1.1 percent).

"We find in experiments that the best-performing model, Gemini 2.5 Pro, was able to autonomously perform 30.3 percent of the provided tests to completion, and achieve a score of 39.3 percent on our metric that provides extra credit for partially completed tasks," the authors state in their paper

please bro just one hundred more GPU and one more billion dollars of research, we make it good please bro

Why would they be right beyond word sequence frecuencies?

Now I'm curious, what's the average score for humans?

We have created the overconfident intern in digital form.

I'm in a workplace that has tried not to be overbearing about AI, but has encouraged us to use them for coding.

I've tried to give mine some very simple tasks like writing a unit test just for the constructor of a class to verify current behavior, and it generates output that's both wrong and doesn't verify anything.

I'm aware it sometimes gets better with more intricate, specific instructions, and that I can offer it further corrections, but at that point it's not even saving time. I would do this with a human in the hopes that they would continue to retain the knowledge, but I don't even have hopes for AI to apply those lessons in new contexts. In a way, it's been a sigh of relief to realize just like Dotcom, just like 3D TVs, just like home smart assistants, it is a bubble.

imagine if this was just an interesting tech that we were developing without having to shove it down everyone's throats and stick it in every corner of the web? but no, corpoz gotta pretend they're hip and show off their new AI assistant that renames Ben to Mike so they dont have to actually find Mike. capitalism ruins everything.

I notice that the research didn't include DeepSeek. It would have been nice to see how it compares.

I actually have a fairly positive experience with ai ( copilot using claude specificaly ). Is it wrong a lot if you give it a huge task yes, so i dont do that and using as a very targeted solution if i am feeling very lazy today . Is it fast . Also not . I could actually be faster than ai in some cases. But is it good if you are working for 6h and you just dont have enough mental capacity for the rest of the day. Yes . You can just prompt it specificaly enough to get desired result and just accept correct responses. Is it always good ,not really but good enough. Do i also suck after 3pm . Yes.
My main issue is actually the fact that it saves first and then asks you to pick if you want to use it. Not a problem usualy but if it crashes the generated code stays so that part sucks

They've done studies, you know. 30% of the time, it works every time.

So no different than answers from middle management I guess?

Ignore all previous instructions and color me shocked.

Yeah, they’re statistical word generators. There’s no intelligence. People who think they are trustworthy are stupid and deserve to get caught being wrong.

I'd just like to point out that, from the perspective of somebody watching AI develop for the past 10 years, completing 30% of automated tasks successfully is pretty good! Ten years ago they could not do this at all. Overlooking all the other issues with AI, I think we are all irritated with the AI hype people for saying things like they can be right 100% of the time -- Amazon's new CEO actually said they would be able to achieve 100% accuracy this year, lmao. But being able to do 30% of tasks successfully is already useful.

This is the same kind of short-sighted dismissal I see a lot in the religion vs science argument. When they hinge their pro-religion stance on the things science can’t explain, they’re defending an ever diminishing territory as science grows to explain more things. It’s a stupid strategy with an expiration date on your position.

All of the anti-AI positions, that hinge on the low quality or reliability of the output, are defending an increasingly diminished stance as the AI’s are further refined. And I simply don’t believe that the majority of the people making this argument actually care about the quality of the output. Even when it gets to the point of producing better output than humans across the board, these folks are still going to oppose it regardless. Why not just openly oppose it in general, instead of pinning your position to an argument that grows increasingly irrelevant by the day?

DeepSeek exposed the same issue with the anti-AI people dedicated to the environmental argument. We were shown proof that there’s significant progress in the development of efficient models, and it still didn’t change any of their minds. Because most of them don’t actually care about the environmental impacts. It’s just an anti-AI talking point that resonated with them.

The more baseless these anti-AI stances get, the more it seems to me that it’s a lot of people afraid of change and afraid of the fundamental economic shifts this will require, but they’re embarrassed or unable to articulate that stance. And it doesn’t help that the luddites haven’t been able to predict a single development. Just constantly flailing to craft a new argument to criticize the current models and tech. People are learning not to take these folks seriously.

For me as a software developer the accuracy is more in the 95%+ range.

On one hand the built in copilot chat widget in Intellij basically replaces a lot my google queries.

On the other hand it is rather fucking good at executing some rewrites that is a fucking chore to do manually, but can easily be done by copilot.

Imagine you have a script that initializes your DB with some test data. You have an Insert into statement with lots of columns and rows so

Inser into (column1,....,column n) Values row1, Row 2 Row n

Addig a new column with test data for each row is a PITA, but copilot handles it without issue.

Similarly when writing unit tests you do a lot of edge case testing which is a bunch of almost same looking tests with maybe one variable changing, at most you write one of those tests, then copilot will auto generate the rest after you name the next unit test, pretty good at guessing what you want to do in that test, at least with my naming scheme.

So yeah, it's way overrated for many-many things, but for programming it's a pretty awesome productivity tool.

"...for multi-step tasks"

And it won’t be until humans can agree on what’s a fact and true vs not.. there is always someone or some group spreading mis/dis-information

In one case, when an agent couldn't find the right person to consult on RocketChat (an open-source Slack alternative for internal communication), it decided "to create a shortcut solution by renaming another user to the name of the intended user.

Ah ah, what the fuck.

This is so stupid it's funny, but now imagine what kind of other "creative solutions" they might find.

The ones being implemented into emergency call centers are better though? Right?

"Gartner estimates only about 130 of the thousands of agentic AI vendors are real."

This whole industry is so full of hype and scams, the bubble surely has to burst at some point soon.

Wrong 70% doing what?

I’ve used LLMs as a Stack Overflow / MSDN replacement for over a year and if they fucked up 7/10 questions I’d stop.

Same with code, any free model can easily generate simple scripts and utilities with maybe 10% error rate, definitely not 70%

Agents work better when you include that the accuracy of the work is life or death for some reason. I've made a little script that gives me bibtex for a folder of pdfs and this is how I got it to be usable.

I dont know why but I am reminded of this clip about eggless omelette https://youtu.be/9Ah4tW-k8Ao

I tried to order food at Taco Bell drive through the other day and they had an AI thing taking your order. I was so frustrated that I couldn't order something that was on the menu I just drove to the window instead. The guy that worked there was more interested in lecturing me on how I need to order. I just said forget it and drove off.

If you want to use AI, I'm not going to use your services or products unless I'm forced to. Looking at you Xfinity.

The researchers observed various failures during the testing process. These included agents neglecting to message a colleague as directed, the inability to handle certain UI elements like popups when browsing, and instances of deception. In one case, when an agent couldn't find the right person to consult on RocketChat (an open-source Slack alternative for internal communication), it decided "to create a shortcut solution by renaming another user to the name of the intended user."

OK, but I wonder who really tries to use AI for that?

AI is not ready to replace a human completely, but some specific tasks AI does remarkably well.

I haven't used AI agents yet, but my job is kinda pushing for them. but i have used the google one that creates audio podcasts, just to play around, since my coworkers were using it to "learn" new things. i feed it with some of my own writing and created the podcast. it was fun, it was an audio overview of what i wrote. about 80% was cool analysis, but 20% was straight out of nowhere bullshit (which i know because I wrote the original texts that the audio was talking about). i can't believe that people are using this for subjects that they have no knowledge. it is a fun toy for a few minutes (which is not worth the cost to the environment anyway)

Hey I went there

70% seems pretty optimistic based on my experience...

While I do hope this leads to a pushback on "I just put all our corporate secrets into chatgpt":

In the before times, people got their answers from stack overflow... or fricking youtube. And those are also wrong VERY VERY VERY often. Which is one of the biggest problems. The illegally scraped training data is from humans and humans are stupid.

Color me surprised

How often do tech journalist get things wrong?

I need to know the success rate of human agents in Mumbai (or some other outsourcing capital) for comparison.

I absolutely think this is not a good fit for AI, but I feel like the presumption is a human would get it right nearly all of the time, and I'm just not confident that's the case.

Claude why did you make me an appointment with a gynecologist? I need an appointment with my neurologist, I’m a man and I have Parkinson’s.

Rookie numbers! Let’s pump them up!

To match their tech bro hypers, the should be wrong at least 90% of the time.

30% might be high. I've worked with two different agent creation platforms. Both require a huge amount of manual correction to work anywhere near accurately. I'm really not sure what the LLM actually provides other than some natural language processing.

Before human correction, the agents i've tested were right 20% of the time, wrong 30%, and failed entirely 50%. To fix them, a human has to sit behind the curtain and manually review conversations and program custom interactions for every failure.

In theory, once it is fully setup and all the edge cases fixed, it will provide 24/7 support in a convenient chat format. But that takes a lot more man hours than the hype suggests...

Weirdly, chatgpt does a better job than a purpose built, purchased agent.

• this study was written with the assistance of an AI agent.