> If the agents become intuitive enough to understand those same inputs directly, what’s left for the human in the middle to add?

Patience. People don't pay others to do work they're willing to do themselves.

Just because they are Agents doesn't mean their output is correct and/or acceptable quality.

> clients already do this kind of integration. They don’t ask their suppliers to "be creative," they give precise goals and constraints.

I'm not sure who your clients are. Mine do have goals, but ask me to be creative and come up with the constraints. Literally, I had someone ask me for exactly that on my last call. My folks are nowhere near precise. If they were, they would not need me.

The person you're referring to is more like a conductor. Everyone in an orchestra can play their instruments just fine. Better than fine. They are great at it. But the conductor keeps them coordinated and synchronized. But these conductors will consolidate and become obsolete too. After everyone else in the chain.

Some ideas question why we would need "middleman" software at all in a decade: what's the point of Photoshop if you can ask AI to fix blemishes and manipulate images and compose your graphics without it? On some level it's selling picks and shovels after the gold rush.

Analogy - If a machinist oversees 10 cnc machines, yes you still need him.

There's always going to be some level of expertise involved and especially with AI some person has to exercise judgement, watching out for hallucinations.

> Why pay for ten salaries—or even one intermediary—when a client could learn to command the agents directly?

Because if it takes one full-time person to coordinate the agents, and those agents together perform a function that the client wants to support some other endeavor, in order to replace the agent coordinator, they would have to abandon the other things they do for the endeavor for which they are hiring the firm supplying the agents.

If the client was literally completely passive and had no other role in their task than hiring the firm for something that wasn't part of a larger effort thet were actively engaged in, then, sure, they would have nothing relevant to the task to sacrifice by being the agent coordinator, they'd just need to spend the time learning to be an agent coordinator, and then actually become an active full-time participant rather than completely passive -- but presumably they were passive because that's exactly the level of engagement they were willing to put in, or, at least, because they found that being more active in the task had some associated disutility.

See the "doorman fallacy".

Customers don't actually want to do the work of their vendors. I don't want to scan my own groceries at a supermarket, or check myself in for a flight (if checking backs, or traveling internationally anyway). If I wanted to be doing that kind of work, I'd get a job at a supermarket or an airline.

I think you don’t really get the idea of engineering. The job is not to spit out code (but even here 10 or even 100 junior-mid developers will not deliver complex enough product without proper guidance from advanced people) - the job is to create a solution. Clients don’t care how exactly it is done, they pay for end result.

Because agents don't have a concept of accountability, even if they're capable enough to replace coordinator, we (humans) probably not going to let them do important things, we always want to have a final say, otherwise how do you know where it is going?

Imagine you are a trader and asked agents to build you a perfect trading system and they discuss every next buy/sell and take a call to execute the action.

How do you know in the long term you are making money? What if they decided to cheat on you and started buying stocks which will definitely crash in 1 month, while showing you a progress of profit by day trading other stocks. Would you just let them do what they're doing?