Most office work never justified its own software
Programs used to need thousands of users to pay for engineers, and office drudge never qualified.
David Senra · Building The World's First AI Software Engineer | Cognition’s Scott WuConsider YouTube. Scott Wu, co-founder of Cognition, which makes the AI software engineer Devin, notes that a great deal of hours went into building YouTube, but far more hours have gone into using it. That is why it passes his test for when software is worth building and what makes it feasible.
His distinction is about cost. As Wu puts it, "if you want to go build a product today, you need a whole team of engineers." Engineers are expensive, with salaries to pay, so Wu argues the math only works out if the software will be used a great deal, "at least like a million times or something," with ten thousand times as his looser second guess.
Wu traces that same job of telling a computer what to do through earlier simplifications. Programming once meant plugging vacuum tubes to make ENIAC do arithmetic, then filling out punch cards, then writing assembly, then writing in BASIC. He says that abstraction will keep climbing, and that you already, in his example, "don't need to know Python or Java" to ask for a website that does certain things, or to ask for a change or new feature in a product you already have.
Most office work never met his usage test. Looking through 15 LinkedIn profiles for certain candidates, filling out forms, or piecing research together into an Excel sheet could all be done with software. But hiring a whole team to build a program you will use once and never again made no sense, so a person just did it by hand.
Wu forecasts that agents will change that by handling the building out of sight. In that future, you give the agent instructions and behind the scenes the agent figures out what code to write and what script will automate part of the task. He describes that setup as "how you control your computer," including for jobs that only need doing one time.
Questions people ask
Will I still need to learn to code to use this?
No. The point of each higher layer, from BASIC to agents, is that you describe the result you want and the system handles the syntax. Knowing how software works can still help you give clearer instructions and check whether the result is right.
How does an agent actually turn words into working software?
You describe a goal in everyday language, such as adding a feature or organizing research. The agent then writes the needed code and scripts and runs them to carry out the task. You see the finished website, spreadsheet, or updated product, not the code itself.
What happens to white-collar jobs if agents do this work?
It shifts from doing repetitive clicks and copying to deciding what should be done and judging the outcome. Someone still has to choose which profiles matter, what the spreadsheet should show, and whether the result is correct. The tedious assembly gets delegated, while the judgment stays human.