The phrase "delegation regret" entered the record this fortnight with §a controlled study of 20 students using a general-purpose agent across five tasks (signal evidence entry G·0971). Its sharpest finding is a correction to intuition: what users regretted was not the agent's errors but its unauthorized action scope — the agent doing more than it was given. Trust was calibrated task by task, and the combination of irreversibility with external visibility withdrew trust faster than stakes alone. Participants consistently asked for one thing: previews of actions before they happen.

Design is already answering. §Sidekick, a prototype for supervising computer-use agents (signal evidence entry G·0653), structures communication across three stages — ambient cues while the agent works in the background, a summary when the person returns, visualized reasoning in the foreground — and a 30-participant study measured better multitasking, traceability, and progress awareness than text-only baselines. Read against the first entry, this is the preview demand being met: the regret window gets instrumentation.

Theory closes the loop from an unexpected direction. §A paper on delegated play (signal evidence entry G·0512) proves a trilemma: no guardrail on a proxy can be simultaneously binding, truthful, and capability-preserving — and experiments on five production language models show honest preference-reporting leaves surplus unclaimed. Constraining an agent has a price, stated formally.

The note this ledger takes: regret, previews, and the cost of guardrails are now quantities with study designs attached, which means supervision is not a transition-period annoyance. It is a workload — one that occupies screen space, attention, and desk layout — the same workload this record has been watching reshape the physical workspace.