AGENTS.md in ADSM: practice and observations
During the development of the Svyazist project I gradually built
the work with the LLM agent through the file structure and textual instructions.
One of the key elements of this practice became the AGENTS.md files.
Over time it became clear that AGENTS.md work not as a set of commands,
but as a mechanism for shaping the context within which the agent makes decisions.
Their hierarchy, placement, and structure began to noticeably affect the stability
and repeatability of the work.
In this publication I recorded my current experience of using AGENTS files: where they really help, what limitations they have, and why their role in ADSM should be perceived as architectural rather than directly controlling.