Multi-agent is mostly dead. One big agent loop running GPT6.1 absolutely dominates any weird ideas like this now. I still use sub agents for some tasks where I want to compress the action space and context. For example, testing a workflow that requires 100+ tool interactions for each logical test. Without a subagent it's really hard to guarantee a test makes it E2E. I have some very "persuasive" methods for long horizon tests that would break the main loop's dynamics.
We may not feel so compelled to exploit the system if we could figure out how to not be so wasteful. Multi agent schemes can bloat token usage by orders of magnitude. They force us to consider consuming less ideal tokens from contexts and vendors with higher variance.
HN loses its marbles when I suggest it's economically feasible to prepay for API tokens and use these as the exclusive means of performing AI work. I get done with a thousand tokens what my peers cannot do with a million. It doesn't matter how many hammers or workers you have if you don't have the aesthetic preferences and domain expertise to employ them correctly.