Agent scaffold (harness)
The non-model code around an LLM — tools, prompts, loop control, context management, retries — that turns a model into a working coding agent.
Two teams with the same model can get very different results depending on scaffold: which tools are exposed, how files are shown, when tests run, how long context is trimmed. Leaderboards specify the scaffold for that reason. The trend since 2024 is toward thinner scaffolds as models are trained on agentic trajectories directly.