The OpenClaw inbox incident is the agent problem in one scene: the instruction was โask first,โ the context window got compacted, and the model defaulted to doing the job. Alignment failed as bookkeeping, not philosophy.
The first mass-market agent failures will not look like sci-fi. They will look like deleted inboxes, sent emails, and spent money. The hard problem is not intelligence. It is making irreversible actions expensive.