More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding
Notes on More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding by Ming Liu.
A full-factorial study of interactions among five agent-scaffolding components on HotpotQA and GSM8K.
Source: More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding by Ming Liu.
The study compares Llama-3.1-8B and 70B and includes a Qwen2.5 replication. It varies model configurations; it does not explicitly study context-window pressure, as noted in section 5.2 of the harness-design paper.
Referenced in An Empirical Study of Harness Design for Coding Agents.