Been thinking about this for a while. The bottleneck for AI agents was never the model. GPT-4 level reasoning has existed for two years. The bottleneck was always: how does an agent pay for things without a human babysitter? And now @kojiru_com is fixing it. Pre-funded wallets are a dressed-up to-do list. The agent stops, you get pinged, you top up, it continues. Models Like this are not autonomy but automation with extra steps. The cold start problem here is real and underappreciated. No credit history means no credit. No credit means no autonomous action at scale. Kojiru’s ACS scoring is built specifically for machines, updating after every task. For me, is also the first serious attempt at solving that loop I’ve seen. Infra before the narrative catches up is usually where the interesting stuff happens.