Title:	Wallenberg WALP project: Sustainable Inference Usage in Agentic Coding Platforms
DNr:	Berzelius-2026-259
Project Type:	LiU Berzelius
Principal Investigator:	Martin Monperrus <monperrus@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2026-09-26 – 2026-12-01
Classification:	10205
Keywords:

Abstract

SuperLean is a research and proof-of-concept project for reducing the inference cost, latency, and energy use of LLM-based coding agents. Agentic coding platforms repeatedly transmit source code, tool outputs, and conversation history to language models; much of this context is irrelevant to the immediate programming task. SuperLean will build and evaluate a transparent middleware layer that applies program-dependence analysis, learned code summarisation, and semantic-diff caching to retain only the context needed for each agent action. The project will quantify the cost–quality trade-off on real software-engineering tasks and develop reproducible methods for deploying efficient coding agents.