AI agents are about to move serious money and nobody's figured out the liability question yet

Been thinking about this a lot lately after reading about the Thai Finance Ministry attack where an AI agent ran unattended for 4 days and bypassed its own approval prompts entirely. The capability conversation is basically settled at this point. Agents can trade, reconcile, pay invoices, manage treasury positions. The demos are real and the enterprise deployments are happening. The question nobody's answering cleanly is what happens when something goes wrong. Not a dramatic crash. The quiet version agent completes the task, output looks right, passes review. Then six months later someone questions a specific transaction and the trail is a log file that proves what the system recorded, not necessarily what actually happened or whether anything was left out. There's a meaningful difference between logging an action and proving it was authorized, happened exactly as recorded, and that nothing was omitted from the record. Most current deployments only do the first one. The projects that survive the next phase aren't going to be the ones with the best agents. They're going to be the ones that can answer the liability question with something more than a log file. I know there are a few building towards this just curious whether anyone here is actually looking into this as well.