I think your general thesis about how AI is used of agents vs. the general chat interface is true, but not the rest.
Your knowledge worker example is where people fail when using AI, and it is usually very obvious.
Knowledge workers who know what they're doing recognized very quickly that AI often goes down very bad rabbit holes, often cites incorrect sources, makes up sources, or will use a random Facebook source or Reddit post as an authority on some topic. In legal circles you have to fight it to understand what venue(s) apply and to get it to stick to the right ones, in medical research it will sometimes pull in crackpot papers from left field and intermix them with well defined research.
People in the knowledge realm caught on very quickly that you have to use AI iteratively and constantly back check what you get and correct it. Let it run on its own and it will give you a legal theory based on some weird case in Kentucky that has no bearing to your issue with a court in Vermont.
Using AI is like using adaptive cruise control: you're still in charge of where the vehicle will ultimately end up and if you decide to go take a nap you will wake up in the hospital or end up in the morgue.
It's true though that the 'time between interventions' is lengthening steadily.
The primary difference is that it's much easier for the AI to figure out itself if it's wrong on the programming side. Knowledge workers don't have automated testing for example.
The end result is the knowledge worker needs to be more heavily involved in the process and be more mindful of mistakes which makes it a worse experience for the user and generally less productive than it is for programmers.
For most professionals it’s just safer to avoid having output in your workflow. Keep using AI, but stop copying and pasting AI output.
In practice, lawyers generally don't ever disbarred except if they lie to clients or steal from them.
All of this will sort itself out and eventually bio, legal, software will all just work.
Don't hold onto any single assumption about the world. Things are going to be weird.
We aren’t talking about the new version of SV spiritualism here, we’re talking about systems people actually care about.
If lawyers and article writers are in the knowledge realm. Then your thesis has plenty of counter-examples of people who just use AI to create stuff and don't correct it.
And I'm probably on the more amateurish end of AI development. As I learned these techniques from YouTube.
When you hear about the work Anthropic researchers are doing, they seem to be using hundreds or thousands of agents to solve problems.
There's a reason they're Anthropic researchers. The value proposition tanks when you actually have to foot the bill for the tokens you use.
Subagent flows can improve overall performance, but it generally also comes with disproportionately high token use. That gets worse with parallel subagents that are often doing the same things at the same time and cannot share a cache.
So the median developer is probably using 5-20 agents at a time for a task. I'm not sure if that crosses the threshold of being a "fleet of agents".
I think that sums up the "fleets of agents" approach fairly well.
I don't actually use AI much at all simply because in my line of work the knowledge is industry specific and niche - an AI can't tell you how to fix a proprietary machine. I have used it a few times but to do dumb things like write some Excel script or some other piece of software I have no interest or need for learning.
It’s a phase.
Fortunately I'm in a shop that isn't being judged by feature output quantity, but is absolutely judged on whether or not our internal or external customers go out of service. We don't have pressure to pump output.
In the sense that "down the stack" is well-captured by existing language, AI is absolutely moving programmers up the stack. I can describe my goals using high-level abstractions, and the low-level implementation is somewhat constrained by that.
How well constrained it is depends on how well I described the problem and the quality of the model - much like how good human-written code is depends on how clearly the authors were thinking when they designed the structure.
I know knowledge workers who throw what you give them into the slop machine and return the slop, but I don't need them, I can feed the slop machine myself. I still need those who have experience and taste that the machine hasn't got yet.