I used to think one big root file was safer. My take now: stuffing everything into the root AGENTS.md is a tax on every call. AGENTS.md is just the instruction file the agent reads. Keeping frontend, backend, release, and security rules in one top-level file feels tidy, but it also forces irrelevant context into work that does not need it.
What changed my mind was seeing the same signal from two places. The OpenAI AGENTS.md guide says instructions are loaded from the project root down to the current directory, so local guidance closer to the task matters. One Cursor community thread reported better results from adding smaller AGENTS.md files in subfolders instead of pushing everything through the root.
That is the only part of tech-leads-club / agent-skills that changed the setup decision for me. This is not just a token-saving trick. It is a context-fit problem. A frontend task should not drag release rules and security boilerplate into every pass just because the repo has one master file. A better filter for updates like this: do not ask whether the tool added more features; ask whether it changes your next setup decision.
Boundary: this is based on the OpenAI AGENTS.md guide plus one community report, not a repo benchmark. But it is enough for a practical next step: keep the root file thin, put rules closer to where the work happens, and share this with anyone still treating one giant root AGENTS.md as the safe default. If your team already split rules by directory, did you draw the line by function, workflow, or risk level?
#AIEngineering #DeveloperTools #TechLeadership #Agentworkflows