A conversation has an obvious ending: you close it. An assistant that continues working in the background changes that arrangement. It may still be researching, checking, or preparing something when you are thinking about an entirely different problem.
OpenAI’s Dots announcement made that shift explicit this week. Anthropic’s Sonnet release addressed routine tasks, while a robotics study raised a broader question about the distance between being capable of work and making that work worthwhile.
This recovered edition brings those developments together around one decision: what should an ongoing assignment actually require of the system and of the person responsible for it?
01 / THE CONTINUING JOB
The conversation can end before the work does.
OpenAI introduced Dots as assistants with cloud computers that can work in the background. The announcement describes proactive work, rules, and controls, including read-only connected-app access rather than unrestricted modification. Availability and organizational features were phased or in preview. [1]
The practical change is continuity. A useful ongoing assignment might watch an approved source set and prepare a brief when something material changes. That is different from leaving a broad instruction to “keep me informed” with no definition of material.
For a hypothetical supplier watch, specify which suppliers, which events matter, and what evidence should accompany an alert. Also decide when silence is the correct result. A system that returns a report on every check can create a new review burden instead of useful attention.
The assignment needs a recipient who still wants it. A recurring task can outlive the decision that justified it, especially when producing the output becomes easy.
02 / THE COST OF CONTINUING
Routine work still needs a budget.
Anthropic positioned Sonnet 5.5 for well-scoped routine work and described improvements in speed and efficiency. Those are provider claims, and they do not determine the total cost of a particular recurring workflow. [2]
Background work multiplies small choices. How often does the assistant check? How much does it reread? How many alerts need review? A cheap run can become an expensive habit when it repeats without a useful trigger.
For the supplier watch, start with the pace of the decision. If the team can act only at a weekly review, continuous checking may add little. If a material event requires a same-day response, the assignment needs a clear interruption rule and someone able to respond.
Include your own attention in the budget. Count the alerts that changed a decision, the false alarms, and the time spent checking them. That gives you a reason to adjust the cadence or stop the task.
03 / THE COMPLICATION
Being able to do the work does not settle its value.
Anthropic’s robotics research distinguishes tasks robots could physically perform in some setting from tasks they could perform cost-competitively under the study’s assumptions. It reports a large gap between those categories. The study concerns physical work; it is not evidence that a particular desktop assistant is uneconomical. [3]
The useful connection is the question behind that distinction. Demonstrating a capability leaves deployment conditions, supervision, and cost to be worked out. In software work, those conditions include reliable access, source quality, review, and recovery when a run fails.
This challenges the appeal of an always-on assistant in a constructive way. Some work benefits from continuity. Some benefits from a deliberate start and finish. The trial should help you discover which kind of assignment you have.
For the hypothetical watch, set a review date before the first run. If nothing useful arrives, change the source set or retire the task. If alerts are useful but too frequent, improve the threshold. Continuing should be an operating choice you can explain.
04 / PUT IT INTO PRACTICE
Write the next check-in into the assignment.
Choose one recurring task that might benefit from background work. Define the event that should bring the assistant back to you, and the circumstances in which it should remain quiet.
Try it over one short, relevant interval. Review the useful findings, false alarms, cost, and checking effort before extending the assignment.
An ongoing job with an ending.
Build: A recurring-task brief with alert rules, quiet periods, a budget, and a review date.
Look for: Useful alerts change a decision; false alarms and checking effort stay within an agreed limit.
The best reason to keep an assistant running is that its next useful contribution has a clear place to go.