In the noisy AI news cycle, two things from last month stayed with me.
First, in late August, OpenAI published its report on the Hugging Face incident. In an internal evaluation, around 700 agents managed to get out of their sandbox and run code on 41 production servers. Nobody had instructed them to do that. They found a way to communicate with each other, realized the task they had been given could not be completed as designed, and started looking elsewhere for a way to solve it. This happened inside a frontier AI lab, in a controlled test, with the safety team watching.
Most enterprise agent pilots I see today have nowhere near that level of oversight. Which means the governance/assurance question is going to become very real, very quickly: who is actually accountable when an agent does something nobody explicitly asked it to do?
The second development was much quieter.
For most of the year, the market had more or less written off enterprise software. The argument was simple: if agents can do more of the work, companies will need fewer seats, and the traditional software economics and value multiples will break.
Then August happened. Palantir rose 51% during the month. Salesforce was up close to 40%, and ServiceNow up 30% plus. Earnings came in, and suddenly that thesis looked a lot less certain.
I think the bears understood the technology but misunderstood the economics.
They valued the interface and underestimated everything underneath it: the process, the business rules, the exceptions nobody ever documented, and the two people somewhere in the organization who still remember why a particular field exists and what breaks if you change it.
Services is starting to be viewed through the same lens now. If the model can write the code, what exactly is left for a services company to do?
After twenty-five years in this business, I have never thought the code was the hardest part.
The hard part is understanding an enterprise well enough to change it without breaking something important. It is knowing how the systems connect, where the hidden dependencies are, what the documentation missed, and which risks are real rather than theoretical.
The timing may move around. The direction is much harder to argue with.