Lita: Light Agent Uncovers the Agentic Coding Capabilities of LLMs

Hankun Dai, Maoquan Wang, Mengnan Qi, Yikai Zhang, Zijian Jin, Yongqiang Yao, Yufan Huang, Shengyu Fu, Elsie Nallipogu · 2025

Read Lita: Light Agent Uncovers the Agentic Coding Capabilities of LLMs by Hankun Dai, Maoquan Wang, Mengnan Qi, Yikai Zhang, Zijian Jin, Yongqiang Yao, Yufan Huang, Shengyu Fu, Elsie Nallipogu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-25873.txt.

Description

Software agents built on language models often come wrapped in elaborate hand-made workflows, which makes it hard to tell what the model itself can do. Hankun Dai and co-authors introduce Lita, a deliberately light agent that keeps only the essentials of autonomy and drops the heavy scaffolding. The point is a fairer, cheaper way to measure the coding ability of new models.

Tested on the Aider Polyglot and SWE-Bench benchmarks, Lita matched or beat more complex workflow and agent baselines while using fewer tokens and far less design effort. The authors propose an agent complexity law: as base models improve, the gap between simple and fancy agent designs shrinks toward nothing. Their results suggest a light agent is enough to reveal what modern models can really do.

Catalog record

CREATORHankun Dai, Maoquan Wang, Mengnan Qi, Yikai Zhang, Zijian Jin, Yongqiang Yao, Yufan Huang, Shengyu Fu, Elsie Nallipogu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-25873

Browse the full library · About the Tower of Babel