Back

Building an Advanced Agentic Harness

14 points1 hourdata4sci.com
budududuroiu28 minutes ago

> The plan is a graph

I much prefer giving the LLM a REPL loop, and injecting all the tools as functions inside the REPL loop.

That means that the LLM isn't constrained to writing a DAG, it can write code that loops, exits early, etc.

hnlqpx99l923 minutes ago

Good stuff, keeping it

dominotw38 minutes ago

why do i hate skillks, harnesses , memory systems whatever. such ideas that everyone thinks they've discovered but are totally useless in practice.

champagnepapi29 minutes ago

agreed. All of these are trying to get to something that can't really every be achieved with LLMs with is determinism. Folks are trying to constrain the models to behave in a certain way with all of these tools, but there's far too many edge cases for them to be reliable. Doesn't mean they can't add some value, but it seems very limited. Hoping it's only a matter of time before we go back to engineering and step back from "vibes".