I think most are actually worth, as agentic harnesses seem to optimize for solving poorly described problems rather than following complex procedures as written. In other words, instruction following maximizing models seem to make worse free-form agents, but they're really all that some domains need.