Connect everything; supervise the parts that matter
MCP as the integration substrate
The published Model Context Protocol provides a common way for an assistant to talk to external tools — a calendar, a project tracker, a customer-record system, a chat application — through a single interface. The practical effect is that an assistant in a workspace can read from and write to most of the tools the worker already uses, without each integration being bespoke. The benefit is breadth; the cost is that the assistant's reach is now wide enough that the boundary between assistance and unattended action becomes consequential. The integration is technical; the discipline is human.
The out-of-the-loop performance problem
The risk that comes with high-reliability automation has been documented in the human-factors literature for nearly thirty years. Operators of highly automated systems show measurable degradation: reduced vigilance, slower detection of unusual situations, and decay of the manual skill needed to intervene when the automation fails. This was not a finding about computers. It came out of aviation, anaesthesiology, and process control, and the empirical record across those fields is unusually consistent. The implication for agentic workflows is that the same dynamic now operates on knowledge work. The more the assistant does well, the less the worker is in the loop to notice when it begins doing badly.
Use, misuse, disuse, abuse
A useful taxonomy from the same literature distinguishes four interactions between humans and automation. Use is the appropriate engagement of the automation on the work it was designed for. Misuse is over-reliance, treating the automation as more capable than it is. Disuse is under-reliance, ignoring useful automation out of distrust. Abuse is deploying automation in settings where its limits are known and unacceptable. All four occur with assistants. The work of designing an honest stance is to identify which interaction is happening in each part of one's working life and to correct the misuse and abuse cases without slipping into disuse.
Where the human stays
The practical rule the literature supports is to keep the human firmly in the steps that involve interpretation and decision, even when the steps that involve gathering and execution are automated. In agentic terms: it is fine for an assistant to gather the calendar, the recent emails, the project status, and the relevant documents into a brief; it is fine for it to execute the resulting actions, like sending the message or filing the ticket. It is dangerous to delegate the judgement in between — what the brief actually means, what the right action is. Most of the failures attributed to automation are failures of having delegated the interpretation step.
Preserving manual practice deliberately
A further finding from the same literature, less often quoted, is that manual skill decays unless deliberately maintained. Pilots who rely on autopilot for years lose the muscle for hand-flying, and the loss only shows itself in the rare cases that demand it. The agentic equivalent is the worker who has not drafted a status note, written a meeting brief, or read a document closely in months because the assistant has been doing all of it. The skill is not lost permanently; it is dulled. A small discipline of doing some of the work manually, especially in the categories the assistant is best at, keeps the worker in shape for the moments when the assistant produces something subtly wrong and someone has to notice.
When was the last time an automated system you relied on produced a quietly wrong result that you did not catch until much later? What had you stopped paying attention to?