miki on Nostr: I have a slight "unease" with the term "agent." Not for the usual reasons, but ...

I have a slight "unease" with the term "agent."

Not for the usual reasons, but because it encourages thinking of agents like we think of employees. Something you have a limited number of, difficult to hire/fire, can only do one task at a time etc.

A better way to think about them would be as (graphs of) ephemeral tasks / jobs, like in Celery, Sidekik and such. In this model, LLMs are essentially an autoscaling pool of compute that those tasks can always tap into.

This way of thinking gives you features like parallel sub-agents, checkpointing, crash recovery or surviving system migrations / power outages for almost free.