Is Context the Bottleneck? Causal Decomposition of Failure in Long-Horizon LLM Agents

Overview

Long-horizon LLM agents degrade as trajectory length increases, but is the bottleneck context quality or reasoning capability? This work introduces an oracle state intervention protocol that causally decomposes agent failures.

Environmental Setup

1. Create a new conda environment

conda create -n agent-decomp python=3.11 -y
conda activate agent-decomp

2. Clone the repository

git clone https://github.com/leqijia/agent-decomp.git
cd agent-decomp
pip install -r requirements.txt