agnostic-ai

Your team uses Claude Code, a colleague prefers Codex, and someone still has Copilot open. Each tool wants its own files, so the same conventions, skills and hooks end up copied in several places, and the copies drift apart without a single error. The tools are not the interesting part here. The workflow is.

We start with what we want from an agent and which part of its setup delivers each outcome: instructions and rules as hints, hooks and permissions as guarantees, skills, agents and MCP servers to get work done. Then we follow that setup through a few tool switches, see where it drifts, and compare two fixes: symlinks, and agnostic-ai, which keeps one spec and writes the native files 25 AI coding tools read. We close with the trade-offs and with questions for the room about what belongs in a shared agent setup.