Technical Context
I'm not focusing on the debate itself but on a gap in the architecture: the orchestrator spawns a subagent, and that subagent drags along the parent model without a native switch. For AI automation, that's bad news because even a simple tool suddenly starts running on an expensive model where a mini-class would suffice.
This is where I usually pause a project and dig into the mechanics. If a tool lacks model selection parameters, no amount of fancy prompt instructions turns into real routing. An agent might 'know' it should save money, but technically there's nothing to choose.
From what's visible in the discussion, the problem isn't the model quality per se, but the absence of a configuration management layer when spawning subagents. That's an important distinction. When model selection lives only at the parent level, the child becomes an expensive copy rather than a specialized worker.
I typically test such things very pragmatically: take a single scenario where the orchestrator decides whether research, summarization, or formatting is needed, and then observe which calls actually hit the API. If every child step goes to a heavy model, tokens burn not because the task is complex, but because the AI integration lacks proper routing.
Searching open materials is honestly not very helpful here: general overviews of LangChain, AutoGen, and CrewAI talk a lot about flexibility but almost never offer a solid answer on how to avoid this inheritance in real production. So, look not at the framework marketing, but at the actual agent initialization code, node-level model binding, and subagent factories.
What This Changes for Business and Automation
The first hit is obvious: budget. If support, research, and validation all run through a single heavy model, pipeline costs climb without benefit.
The second hit is quieter but more painful: you lose predictability. You can't properly calculate unit economics if subagents can't switch between cheap and powerful modes based on the task.
Those who build AI solutions for business with explicit model control at every node will win. Teams that trust the orchestrator to 'somehow figure it out' and optimize cost without parameters will lose.
At Nahornyi AI Lab, we clean up exactly these bottlenecks: we decompose orchestration by roles, implement explicit model routing, and remove hidden token guzzlers. If you already have an AI implementation but your bills are growing faster than the benefits, let's examine the call graph and build an AI automation where the system thinks where it matters, not in every single breath.