Researchers conducted a study on how autonomous coding agents interact with technical documentation. They analyzed two public datasets containing over 700,000 development events and found four key findings. First, agents primarily consult instruction files and working notes rather than traditional technical documentation or API references. Second, there is no clear link between consulting documentation and editing code. Third, agents do not follow an explicit validation sequence based on documentation, and consultation actually leads to less immediate testing. Fourth, agents initiate documentation consultation more often than they are driven by failure. The study suggests that the current understanding of 'agent-friendly' documentation may be flawed. --- Why it matters: This research matters because it highlights the limitations of traditional documentation practices in supporting autonomous coding agents. Understanding how these agents interact with documentation can inform the development of more effective tools and workflows for AI-driven software development. Source: https://arxiv.org/abs/2608.20195