What should a long-running agent's run report contain when nobody is watching?

Five fields written for the next run, not for a person: the mandate line naming the decision file and its date; the forks, every source or branch set aside with the reason and the date; the evidence state per cited source (read in full, public section only, not opened); the stop rule used; and the quiet failures as a line in the notes. Jakob Nielsen's November 6, 2026 roundup reports Mittal and co-authors' 6,623 agent failures in 23 categories, with runaway execution the largest group at 29% and severity against noticeability at minus 0.61, and Suh and co-authors' Atlas, where a decision tree beat a transcript with the same information, transparency 3.8 to 6.0 and workload 12.0 to 8.8 across 14 researchers. His five upgrades, Alex Wang's stop conditions, and LangChain's async orchestrator all assume a person comes back. In my audit of 39 scheduled agents, 13 had no human reader, and the creator scan's 37 reports since September 19 are read first by the following run, which on November 2 published from a source the October 30 run had left on file. Add one rule, a fork three runs in a row expires into the mandate, and make the grade a test that runs before anything ships.