The number one bottleneck is by very far the shit sloppy-pasta code written in LLMish that models produce. Arguably agents are way more advanced tech than LLMs.
When the thing has to verify that an input parameter, passed as an integer, is between 0x000a0000 and 0x000affff and the model decides to pull a library and converter the integer to a string and then do substring matching on "0x000a....", you kinda get an intuitive feel as to how these thing do "reason".
Don't get me wrong: I pay three AI subscriptions and use models daily to help me.
Which is why I know the bottleneck is the shit code the models do produce.
> but I am also wondering if agents would be more useful if they had access to the same infrastructure, software and environments we do as human developers.
I don't know if you're been living in a cave or something: for they do have access to just that.
The number one bottleneck is by very far the shit sloppy-pasta code written in LLMish that models produce. Arguably agents are way more advanced tech than LLMs.
When the thing has to verify that an input parameter, passed as an integer, is between 0x000a0000 and 0x000affff and the model decides to pull a library and converter the integer to a string and then do substring matching on "0x000a....", you kinda get an intuitive feel as to how these thing do "reason".
Don't get me wrong: I pay three AI subscriptions and use models daily to help me.
Which is why I know the bottleneck is the shit code the models do produce.
> but I am also wondering if agents would be more useful if they had access to the same infrastructure, software and environments we do as human developers.
I don't know if you're been living in a cave or something: for they do have access to just that.