I’m hardly the first person to posit that the “AI” industry is in a bubble and due for a crash. I doubt I’m the only person who hopes it takes out OpenAI and Anthropic1 in the process. But if this bubble bursting takes out anything in its blast radius, then please, for the love of all that is good and holy, let the agents have to “dodge this.” LLMs gave us queryable documentation. The linear algebra behind them has given us better full text search. We’ve gotten some really amazing benefits from the “AI” boom. Agents, on the other hand, aren’t it, and it’s time we stop treating them as magic, people, or as the go-to answer for everything.
Oh sure, the premise of all these agents sounded amazing. Finally, a program that can do things for you without relying on a bunch of cobbled-together if-statement edge case handling and a prayer. Except…all it’s done is get expensive. Well…there’s also the bad behavior in production, and that letting agents run wild is a bit of a security nightmare. But what we’re starting to realize is, that’s only as good as the percentage of time it’s completely right, which is disappointingly rare. Now, the faster these things give you feedback to address, ideally with the smallest set of changes (if any), the less annoyed you likely are when they’re wrong. Getting a small code snippet that you could immediately test is less annoying than letting it run for several minutes, try to 1-shot a feature, and be wrong somewhere. The problem is the “AI” hype bros are trying to push us in the opposite direction – “agentic engineered loops that can run for days without you, man.”
Enough with that crap. Agents demoed well, and were a fun novelty, but they need to to crash out and the economics of LLMs get aboveboard, instead of relying on massive subsidies from the big names. Between the price hike (on top of already increasingly expensive agents) and the slowly dawning realization that they hype and reality are far, far from each other, may convince us to push LLM usage into more productive directions.
First, and sadly, I don’t think agents go away completely. I do hope that they’ll turn into the LLM equivalent of reflection in programming languages – incredibly powerful, equally as easy to shoot yourself in the foot as it is powerful, and if you’re going to reach for it you should be prepared to justify why you need it and why no other (non-agent) option works in your case. Yes, that means you can’t just be a “loop engineer” and have to do real programming, but it’s high time we got back to that anyway.
The fundamental problem with agents is the thing that creates the hype behind them – agency. Namely, the more agents exercise, the less we have. Look back on the Internet’s history of ens***ification, all for more money than it’d cost to just pay to have humans do it the way we did pre-2022 (and do it better). Probably the only good thing that’s come out of “agentic engineering” was the work the relative handful of companies actually doing it well put into building their testing infrastructure to keep the agents from completely f***ing everything up. I’m not talking about a line in Agents.md telling it to make sure you have unit test coverage – these guys built tools to run their application in isolated, heavily monitored sandboxes so they can validate every aspect of real behavior in real running software, which is stuff we all should have been doing for years.
And to any of the AI hype bros reading this who are starting to say “the agent can do it better and faster than I can, so I want them to be able to make the decisions instead” – either you suck at development, or your definition of “better” is “more code faster,” (I arguably repeated myself there). In that case, yes, a hard reversion to the mean is an improvement over you. The implication of that is not that the agents are “intelligent.”
Ironically enough, as I was writing this, this video came up in my queue, with a great set of case studies where agents took something that, when done manually, would have been cheaper and the problems encountered would have been much cheaper before being solved. Instead, they blew up whole budgets.
It’s not just the runaway costs in places that were better off being done by hand. For all the hype about how agents are “solving” software engineering, most companies aren’t shipping more. Out of the companies heavily pursuing AI usage internally, most of the bigger names have a sinking reputation in terms of software quality. AWS pushed heavy AI usage and immediately started self-owning, Microsoft’s Github started suffering so much they just changed how they report uptime (and the Internet had to re-create it), and Spotify brags about deploying thousands of lines of code per day with users seeing no improvement in experience…at best.
This is the part where people generally start waving their hands around, talking about how you still need humans in the loop to supervise the agents. That line would be a lot more convincing if it weren’t coming from the same people trying to push humans out of the loop altogether (because we’re the bottleneck). LLMs have potential as useful tools – the problem with agents is that they rely on the assumption that the underlying LLMs can be more than what they are.
In short, the failings of agents is the hyped-up claim of “Artificial General Intelligence (AGI)” hitting the reality that the underlying technology powering the “magic” is nothing more than the result of a lexigraphical popularity contest. If agents were just an experiment interested people were trying out, this would be no big deal. Maybe we’d find a few places where it’s handy and have them in our back pockets for these niches. Instead, we were so obsessed on whether we could get these things to make decisions, we didn’t stop to consider if we should. Now, they’re forced on us, and we’re all trying to clean up the messes made when agents fail to clear the gap between hype and actual capability.
IBM called this over 40 freaking years ago:
A computer can never be held accountable.
Therefore, a computer must never make a management decision.
We’re trying make that word “management” do a lot more work than it should. We think that so long as agents aren’t making decisions we reserve to people who’s job title is literally “manager,” some wrapper around fancy autocompletes can do whatever said fancy autocomplete spits out. But that’s not the point. The point is that computers shouldn’t be making any value judgements, and the apparent point of agents is to make as many of them as possible right up to (and sometimes including) the job of the person pushing them.
We got some useful stuff out of LLMs, and that’s great. Agents make for cool demos, but are showing a lot less value in practice, There may well be some redeeming value there, but “AI agents” are not a catch-all solution to everything. Once we accept that, and purge the agents and then see where removing them actually lowered our quality of life, then we can start seeing some consistent value from these tools.
- I’m not naive enough to think the whole industry will go under, never to return, but taking out the “big” players would certainly be wonderful. ↩︎