I think orchestration makes sense. I don't want to manage twenty conversations just to get a few things done. If the agents can work on their own, I shouldn't need to keep switching between threads to find out what happened and tell each one what to do next.
The useful part is having one place to talk. I give it something to work on, it delegates, and the results come back there. If something needs a decision, it asks me. I can change direction without having to repeat myself in every session.
That's different from saying everything needs five layers of managers.
One conversation, a few different choices
There are a few things getting grouped together under orchestration. One is the interface I use. Another is how many agents do the work. Another is how many other agents a result passes through before it reaches me. Those aren't the same choice. A hundred workers could report to one coordinator. A smaller system could still have a long chain of handoffs.
The conversational part is showing up in consumer products too. Meta's Muse design write-up describes a main conversation, with side chats for people who want separate context. OpenAI describes dots as handling several projects without making you juggle separate threads. [S2, S3]
That makes sense to me. It also doesn't tell us how complicated the system underneath needs to be.
What gets lost on the way back?
My concern with deeper orchestration is what makes it through all those handoffs. A worker finds something, another agent summarizes it, and another decides what I need to see. By the time I get the answer, can I still get back to the original result? Can the agent I'm talking to check it, or is it relying on someone else's summary?
I don't think every summary is a problem. Summarizing is part of why this is useful. I don't want every intermediate step dumped into the main conversation either. But there's a difference between not having to read the details and not being able to inspect them.
The product designs acknowledge some of that distinction. Muse has an activity log, and OpenAI says you can open your dot's computer to inspect its work. [S2, S3] Those features don't prove the systems always get things right. They do show that a simple conversation doesn't have to be the only view of what happened.
More agents can still make sense
I can see the argument for a large swarm when the job is to try a lot of possibilities. If the work can be split into independent attempts, doing more of them at once could be useful. That's a different situation from passing the same task through another manager.
There's also a reasonable argument for another layer if it has a specific job, like checking results or coordinating a genuinely separate part of a project. I don't think fewer layers should become a rule that gets in the way of work that actually needs them.
What I'm less convinced about is needing that scale for most ordinary tasks. Having more agents involved isn't, by itself, a reason to expect a better result. I'd want to know what they're adding and whether I can still follow how the answer was reached.
Where I'm at now
I already use orchestration. I have a supervise skill adapted for my own remote setup. This isn't me arguing that we should go back to manually steering every session.
I just think the useful interface and the size of the system are separate questions. I want one place to delegate, see what came back and make the decisions that need me. I'm not convinced that requires a long chain of agents managing other agents.
Maybe much larger systems become practical as the tools improve. I'm open to that. For now, I don't have a number of agents that I think is always right. I want to understand why the next one needs to be there.