scaffold

This paper shows that the choice of agent harness (scaffold) can cause up to a 40x difference in tokens per solved task, while model pass rates vary only slightly, demonstrating that harness–model pairs, not model alone, should be compared for human-centered coding-agent evaluation.

A survey of 239 papers analyzing how AI agents self-improve by updating the model itself or the scaffold (prompts, memory, tools).

This paper presents a novel blueprint for self-improving agents that combines scaffold editing and weight training through a meta-agent and feedback-agent, achieving a 14x speedup on a CUDA kernel for AlphaFold.

This article describes a scaffold that scales test-time compute on Qwen-3.6-27B and Gemma-4-31B using iterative corrections and branch exploration to surpass Claude Mythos in code optimization. It includes a paper link and GitHub repository.

The author proposes that durable memory in multi-agent AI systems should reside with the project owner, not with task specialists, drawing inspiration from consulting firm practices and academic literature on project memory. They present a research scaffold with templates and evaluation rubrics for a field trial.