This is the direction the whole ecosystem is moving — declarative agent definitions over imperative code. The spec file approach has the same advantage CLAUDE.md has for Claude Code: it separates WHAT the agent should do from HOW the runtime executes it. That means you can version control your agent behavior, diff changes in PRs, and roll back bad deployments just like any other config. The agents that win long-term will be the ones where the spec is human-readable enough that non-engineers can audit what the agent is actually allowed to do.