Tag
This paper investigates the minimal context needed for coding agents to edit code, finding that natural-language summaries of code are ineffective and that surrounding context matters little, with compressed context achieving equal results at a third of the tokens. It also reveals a noise floor due to temperature-0 API inference.