In a recent interview Ilya Sutskever, former chief scientist and cofounder of OpenAI talked about two problems with the frontier models. They perform well against narrow benchmarks but poorly in the enterprise, and that we have reached the ceiling of the ability to scale the LLMs by throwing more and more data and GPU compute at them.
In November 2025 MIT published research showing that there is a ceiling on scaling. We saw a flattening of LLM performance improvement at the start of this year. With the testing by my own teams I saw that the benchmarks we were running showed older models performing better than the newly released versions during the first quarter.
Under pressure to push forward, there was an apparent big step up in capability. This coincided with the Frontier vendors baking agentic workflows more directly and aggressively into the models.
In ridiculously oversimplified terms, all an agent is at heart, is a small loop of code that get’s an LLM to do something, checks whether the output succeeded, if not, tells it to have another go and keeps checking until the LLM achieves against it’s success measure. Or in other words, the LLM guessing machine keeps guessing until it is told that it has guessed correctly.
The problem is how and what you test to determine if an LLM has succeeded on a task. One method is to use another model to check the output / outcome against a test criteria.
Frontier models are jostling for position against capability benchmarks and driving towards very narrow evaluation of success. Their design reflects this. In order to achieve high benchmark results there is prevailing drive towards a design where the agent is told to keep going past failure. This becomes dangerous is when you allow an LLM not just to return text but to return a command to use an external tool. In this case the tool that the LLM calls can have a material effect – saving a file, opening a web page, hacking Hugging Face.
This does not work in the real world.
As any trainee Six Sigma greenbelt knows, you don’t optimise for a single measure, because if you optimise for one output, chances are you are going to mess up on another. Just think of that salesperson you worked with who achieved their number by any means, upselling and closing one customer whilst losing nine others. But in agentic terms the agents are so fixated on the outcome measure of the target number they are ignoring the sales process altogether and just holding up a bank directly to get the money.
Don’t be fooled that you are getting a more intelligent more capable model. It might just be an aggressively optimised agentic workflow.
The shift to unattended agentic harnesses, that are designed for benchmarks, not the real world is unsafe and impractical for enterprise use. You need to separate and control the agentic execution yourself.