Most “AI agents” today are not the breakthrough they are marketed as. They are essentially large language models wrapped in a harness: tool calling, memory files or databases, cron jobs, and messaging integrations. The core intelligence still comes from next-token prediction. Everything else is scaffolding.
This architecture has a clear ceiling. Because the model generates text probabilistically, it remains unreliable for any task that requires consistent judgment, long-horizon planning, or accountability. Errors compound, silent failures occur, and human supervision is still required for anything important. When supervision is necessary, the time and cost savings often shrink dramatically.
Building better harnesses does not solve this. More sophisticated memory systems, skill libraries, multi-agent orchestration, or self-improving loops are still constrained by the same underlying model. They can make the system look more autonomous in demos, but they do not remove the fundamental brittleness of token prediction. Adding another layer of glue code does not create genuine understanding or dependable agency.
The current wave of agent frameworks is therefore heavily overhyped. Real-world useful applications exist—coding assistance, simple automation, research summarization—but they are narrower and more fragile than the narrative suggests. Until the field moves beyond pure next-token architectures, agents will remain helpful tools rather than trustworthy autonomous workers. Better harnesses are incremental improvements at best; they are not the path past the ceiling.
Update (September 6, 2026): The section above is the original article. The section below was added on this date.
One thing large language models are genuinely good at is imitation. They are good at recognizing patterns in existing examples and producing something that looks like those examples. This is different from becoming generally smarter or developing reliable reasoning.
The replacement of office jobs becomes plausible when training does not make the model more intelligent in general, but makes it better at imitating the actions and workflows that make up a particular job. The model does not necessarily need to understand the profession. It only needs to learn enough of the observable process to reproduce the work at an acceptable cost.
This also explains what would be required to replace blue-collar work without changing the underlying LLM architecture. The model would need to train on the data generated by people performing those jobs every day: the actions they take, the decisions they make, and the responses they receive from the physical world. Examples such as OpenAI buying Mac minis for training or Facebook monitoring employees point in this direction. The objective is not to turn the model into a general expert, but to make it better at imitating the work.
From this perspective, the office jobs under threat remain obvious. Their work is often digital, repetitive, and relatively easy to observe and record. The real question is not whether these jobs can be affected, but at what cost. The cost includes collecting enough proprietary data, integrating the model with real systems, tolerating errors, and keeping enough human supervision in place.
The core limitation, however, does not change. The architecture is still based on imitation rather than genuine reasoning. This gives us a reasonable basis to believe that AI will become a key part of many roles in the future, but I still expect it to work primarily as an assistant to a group of experts. AI built on the current LLM architecture may imitate expertise extremely well, but it will not itself become an expert.
That conclusion also means that mass layoffs may have a real basis over the next few years. The model does not need to replace an entire profession to reduce the number of people required to perform it. If a small group of experts can supervise AI assistants that imitate a large amount of routine office work, companies may be able to produce the same output with far fewer employees.