The LLM-and-tools loop behind a simple AI agent
A reply describes an agent that thinks, calls a tool, observes the result and repeats. It identifies ReAct, from October 2022, as the earliest key formalization of that approach.
TLDR
The reply says ReAct interleaved reasoning traces with actions, such as Wikipedia searches, to tackle multi-hop questions and fact checks. It describes BabyAGI, from March 2023, as an approximately 140-line demo that generated, prioritized and executed tasks using GPT and vector memory. Early runs often looped or stalled.
Combined views
613
1 Source, first seen 10h ago