OpenAI is cutting off Cursor. Your AI product may depend on someone else’s model.
Welcome back.
One number in OpenAI’s latest research report deserves more attention than the rest.
Inside its research organization, OpenAI now uses 3.1 agent workdays for every human workday.
Researchers are running several coding agents at once. Those agents write code, troubleshoot infrastructure, monitor experiments, and take on assignments that would occupy a skilled researcher for hours or even days.
OpenAI says this has helped its researchers produce more code and run more experiments. It also says its systems have reached what it calls an automated research intern: an agent capable of completing clearly defined research tasks under human direction.
This is internal data from an AI company doing highly technical work. It does not mean every employee can suddenly manage three digital workers. OpenAI also notes that more compute contributed to the increase in experiments, so agent use cannot claim all the credit.
But it does show where work is heading.
The important AI skill is shifting from getting one good answer to directing several streams of machine work.
In other words, your strongest employees may soon operate less like individual contributors and more like managers of AI labor.
One person is becoming a small team
Most companies still teach AI as a conversation.
Write a prompt. Receive an answer. Improve the prompt. Try again.
OpenAI’s researchers are working differently. They can give separate assignments to several agents, let those agents work simultaneously, and step in when the work needs direction.
That creates more capacity, but it also creates more work to manage.
OpenAI found that more than half of its successful tasks estimated to take a person four to eight hours still required at least one human intervention. The agents may do much of the execution, but people still set priorities, judge results, and decide what happens next.
That distinction matters.
An agent is not valuable because it can work alone. It is valuable when a person can give it a bounded assignment, inspect its reasoning, correct its direction, and confidently use the result.
Imagine an operations leader preparing a weekly performance review.
One agent cleans the data. Another investigates unusual changes. A third compares performance with the previous month. A fourth drafts the briefing.
That sounds powerful. It can also produce four polished answers built on different assumptions.
The leader still needs to define the question, choose trusted sources, resolve conflicts, and own the final decision. More AI output without that structure simply creates more material to check.
Give agents assignments, not prompts
If you want one person to manage several agents, every assignment needs four things:
An outcome: What should exist when the work is complete?
Boundaries: What may the agent access, change, or decide?
Evidence: What sources or calculations must support the result?
A review point: When must a person inspect or approve the work?
This is closer to writing a good brief for an employee than writing a clever prompt.
It also changes how leaders should measure AI adoption. The number of prompts sent tells you very little. A better set of questions is:
How much useful work did the agent complete?
How often did a person need to intervene?
How much correction was required at the end?
Did the completed work improve speed, cost, quality, or growth?
The goal is not maximum autonomy. The goal is dependable leverage.
Try this with one real workflow
Choose one recurring piece of work that takes several hours and can be divided into clear parts.
Do not hand the whole process to one agent. Break it into separate assignments. Define the inputs, expected output, and review point for each one. Then run the work in parallel while one person remains accountable for the final result.
Track what came back usable, what needed intervention, and what created additional review work.
You may discover that the agents save hours. You may also discover that coordinating them becomes the new bottleneck. Both findings are useful.
OpenAI’s report is not telling every company to copy the way an AI lab operates. It is showing us that the unit of work is beginning to change.
Soon, asking whether an employee uses AI will be far too simple.
The more useful question will be: How much machine work can this person direct without losing quality, context, or control?
Haroon
What recurring workflow would you trust one person to run with several agents? Reply and tell me.