what's the deal with agents?

i mostly managed to ignore the ai agents hype train that was garnering speed throughout 2024. the applications side rarely hooks my interest, and that's fine, there's plenty of stuff to keep busy with. but if it was difficult to avoid agents in 2024, i imagine it will be next to impossible in 2025. i think this might be the year of ai agents. so to commemorate the year of agents i want to dip my toes into this space and just get a feel for what's going on, where are we at? to help me out i've been blessed with a livestream from hu-po, where he covers just what i'm looking for.

the definition of agents are fuzzy, ....

agent frameworks

langgraph. seems to be the most popular alternative out there. it's been around for a while and it's probably the most extensive framework out there which also becomes one of its pitfalls if your just looking to get your hands dirty. naturally this builds in to the whole langchain ecosystem...

pydanticai. a cleaner alternative to langgraph that, while not as full-fledged, is easy to get into and start using.

smolagents. recent release i've heard a lot about from people in my circles. smolagents seamlessly incorporates HF models, but it also comes with a huge baggage in transformers. smolagents gives the agent access to a python interpreter which it can use freely to do things like tool calls. this works a lot better than json because python is a natural abstraction layer.

browser use. instead of an agent with tool usage that is constrained to a number of alternatives through json tool calls. this library enables an agent to freely use your browser through a visual lm. this removes the dependency on internal tools like calculators or api calls etc.

future

expect heavy churn in this sphere. there is an insane amount of frameworks and libraries out there that basically achieve the same thing. don't worry too much about becoming irrelevant if the library you chose to learn stops being used, the core abstractions here are the same. the goal is skill acquisition and familiarity with key concepts. a thread of core abstractions persists.

just like the insane growth of the cloud and cloud tools created a new role called devops, the growth of ai agents may create a new job of agent ops for people tasked with understanding the libraries, abstractions and how to glue agents together to make them to what a company wants them to do. these agents take a lot of configuration

[wip]