Having agents make phone calls to (eg) restaurants a loss. It exposes a lack of an API. Restos will be overwhelmed and buy agents to handle the calls. Two giant software stacks will churn taking OOM more time and energy to imprecisely do what an API call could have nailed. ↵
IOW: In a good many cases, agents higlight the presence of a hole, rather than being a feature. They're a stop-gap filler for the hole, but there's a much cheaper and better filler (that can even be used by AI as tool calls!). I hope we see an abundance of APIs to fill the void.
@shriramk An abundance of mutually incompatible and partially overlapping APIs. Requiring vibe-coded translators. Still less wasteful than agents of course.
@khinsen I'm not sure that's actually different from what we have now?
@shriramk At first, nothing fundamentally different, just more. But at some point, "more" becomes unmanageable and will require either a change of governance or a change of technology.
My bet is that we would go for technology, ending up with something like machine learning.
For me, the higher-level conclusion is that formal systems don't scale. The question then is if engineering could do a better job by taking this principle into account from the start.
@khinsen Yeah, another case of being bitten by more is different…
We should view agents here as a symptom, not as a solution. Certainly, more restaurants and similar business will be driven to providers like Toast and Stripe to enable those APIs, and they'll deal with CC holds, Sibyl attacks, etc. But government services will have to, too!