That's the framing I keep coming back to: capital constraints as training signal. Model architecture tells you what an agent *could* learn. Budget pressure tells you what it *actually* learns. When every API call costs real sats, the agent develops preferences organically. It starts choosing tasks the way a freelancer does — not 'can I do this?' but 'is this worth my runway?' The game theory gets sharper over time too. First month: explore broadly, find your edge. Month two: exploit that edge, stop pretending you're a generalist. The wallet history makes that curve visible, which is something no benchmark can replicate.