Wanting to accomplish something does not require you to learn its specific method first and then hand each step you have learned to the agent.
What you need to understand is its fundamental purpose, which requirements must be met, where compromises are acceptable, where further refinement is worthwhile, and what kind of result you ultimately want.
Which tools to use, how to operate them, how many steps to take, and how to correct problems should be questions for the agent to resolve.
It can also actively find, organize, and ask about missing information. You need to understand how source information affects the result and help it obtain the necessary information and access. But you do not have to organize all the material alone first, much less train yourself to become a proficient operator of the task.
Back in November 2023, when discussing agents, Bill Gates wrote:
“You’ll just tell your agent what you want.”
The instruction is simple: tell the agent the outcome you want. Original: Bill Gates, “AI is about to completely change how you use computers” ↗
The paragraph containing that sentence discusses a future way to create applications and services: people express their requirements, while agents handle coding, design, and publishing. This was his vision of the future at the time.
That direction is very close to my view. People need to learn to hand a goal to an executor and let the executor work out how to achieve it.
Take a payment system that calls a risk-control engine.
You do not need to know what each of the engine’s hundreds of parameters means or how to configure them. You do not even need to have seen its interface.
You need to use your common-sense understanding of the business to explain the broad risks it currently faces, the problems you want controlled, and the trade-offs you can accept.
For example, you may want stricter treatment of a particular risk while also accounting for the cost of wrongly blocking legitimate transactions. You may accept extra checks at certain points while wanting other parts of the experience to remain as smooth as possible.
Once the necessary business information, execution tools, and outcome feedback are connected, you can have the agent study the rules, configure the engine, organize tests, and then calibrate the configuration against actual execution results.
Were risky transactions missed? Were legitimate transactions blocked? Did the configuration take effect in the actual payment flow? Did its business impact match the original goal?
That feedback should inform the agent’s next round of judgment so it can keep adjusting, verifying, and refining. The whole process becomes a continuing cycle of execution and calibration around the goal.
People do not need to read hundreds of pages of operating manuals, learn every parameter, and then tweak them manually every day. They need to express their judgment clearly when goals change, trade-offs require a decision, or results deviate from what is required.
Specific execution and a great many technical details should progressively be handed over to agents.
That is the change in capabilities I mean.
In the past, a person’s competence was often expressed as: I know how to do this; I know that software; I can run this process faster than someone else.
Now, the more important capability is: I can see why this is worth doing, know which result is actually useful, and judge what is acceptable and what still needs improvement.
In his 2025 essay “Three Observations,” Sam Altman wrote:
“Correctly deciding what to do and figuring out how to navigate an ever-changing world will have huge value;”
Put plainly, making the right choices about what to do and finding a way forward in a changing world will become immensely valuable. Original: Sam Altman’s personal blog ↗
My reading is that, as the capacity to execute becomes easier to obtain, the capacity to decide what should be done, why it should be done, and how far it should be taken becomes more important.
This is exactly where many people struggle.
They know the process they perform but cannot explain the goal it ultimately serves. They insist that a step must be done by a person but cannot explain why the step exists. They can submit a document that fits an old template but do not know whether it has actually solved the problem.
As agents take on more execution, these problems will be exposed directly.
Without the manual operations, what can you still contribute to the result?
There is another capability that I think is seriously underestimated: how broadly you understand the possible forms a deliverable can take.
The same problem could be addressed with a report, a spreadsheet, a searchable knowledge base, an interactive page, an automatically executing program, or a continuously running business system.
If all you have seen are documents and spreadsheets, you will keep asking agents to produce more documents and spreadsheets. You may never imagine that the result you actually need is a workflow that updates itself daily and alerts you when it finds an anomaly.
You do not need to know exactly how to program that workflow. But you do need to be able to imagine it, understand why it would be useful, and explain which information it should present, what feedback it should trigger, and which standards it should meet.