Clara Shih left Meta after her own agents made her stop posting entry-level jobs

Where each witness sits explains most of the disagreement. Aaron Levie sells software and Matt Garman sells the compute underneath it ^[2][3], and a forecast that technology creates more jobs than it removes is also a forecast that their customers keep buying tools for people. Eugenia Kuyda and Amjad Masad are running the substitution: one has stopped buying junior engineering labour, the other has stopped buying big enterprise software contracts and writes the code in-house instead ^[4][5]. The sceptical camp is not purely commercial, since Kathryn Ann Edwards is a labour economist rather than a vendor ^[3]. But only one group here is reporting on its own payroll.

Clara Shih's account is useful because it describes a process rather than a person. A product development cycle that once needed user researchers, designers, product managers and three kinds of engineers came down to one or two people and a prototype, with comparable gains in marketing, distribution and privacy review ^[6]. She made the change quietly, without any restructuring announcement: she stopped renewing entry-level postings because she no longer felt she needed them ^[7], and she did that in the autumn, roughly two quarters before she actually left the company in the spring ^[3]. Requisitions move before policy does, which is why headcount forecasts built from announcements lag the thing they are trying to measure.

The numbers deserve to be held apart. Shih predicts one in five corporate roles is "especially going to be challenged" ^[11], which is 20 percent ^[1]. Platformer notes Katie Paul's reporting for Reuters that Mark Zuckerberg wanted to cut as much as 60 percent of Meta this year on AI efficiencies, and that the plan was derailed partly by agents that underperformed ^[9]. Shih's figure is about a third of that ambition ^[2], and it is a claim about a decade rather than a fiscal year. Both can hold: agents good enough to make a hiring manager close a junior req are not agents good enough to remove three-fifths of a company.

Two facts complicate Shih's account. She now runs a nonprofit and a consumer brand whose customers are entry-level workers ^[10], so her incentives point toward alarm, and Meta's own experience is evidence against her thesis ^[9]. Her most load-bearing statement is the one that cuts against her old commercial interest, not her new one. She built and sold agent platforms at Salesforce and then at Meta ^[8], and she now says the story she told about them, that automating rote work frees support staff for higher-order tasks, has "primarily not been true" ^[12].

Waiting for cleaner data feels prudent when credible people disagree, but the two errors here are not symmetric. Freeze entry-level hiring on the deployers' read and the optimists turn out right, and you rehire in two years at a premium into a market that has not thinned. Keep the ladder and the deployers turn out right, and you carry visible cost for several quarters while your competitors do not. Most of the builders in this series took the comfortable side of that bet by saying jobs would be different rather than fewer ^[13], and comfort is the part worth discounting. What an operator actually holds this quarter is an internal record, not a forecast: which junior requisitions managers stopped reopening in the months after an agent shipped, and whether anyone was asked to explain why.