סוכנו שלא נכשל, אלא החיבור שלו

תקציר מקורי באנגליתTwo runs touch the same session, the second write silently erases the first, and the agent keeps answering with total confidence from stale state. Nothing crashed and the model did not hallucinate, so this is a harness failure, the kind that lives in the system around the model rather than in the weights. Using OpenClaw as a public case study, Vinoth Govindarajan walks the usual suspects: state that was never persisted, overlapping writers with no single writer lane, a tool call that never returns because nothing set a deadline, and an approval that outlived the action it was supposed to authorize. The through line is that a model only proposes; the harness has to commit, and a receipt has to prove it. A transcript shows what the agent said, but a receipt is the evidence that survives: it