Zuckerberg admitted during the all-hands meeting on July 2 that the progress of AI Agents "has not accelerated as expected" over the past four months.

Meta spent $145 billion on AI infrastructure this year, but the agents couldn't be built, leaving 35% of computing power idle. They had to sell computing power to others instead. This caused Micron and SanDisk to plummet, dragging down global chip stocks.
Current AI hasn't reached AGI yet. If you ask it to plan, execute, and correct errors on its own, it can't. Claude Code writes code, Cursor modifies code, and Grok searches for information—these are already producing real value because humans are driving and AI is assisting.
But the Agent logic is the opposite—AI drives, humans watch. This requires not bigger models, but stronger reasoning, planning, and self-correction capabilities. Current models are far from that.
Meta spent $145 billion and still couldn't make progress, which shows one thing: Agents can't be built just by throwing money at them; they require breakthroughs in underlying capabilities.
When AGI arrives, Agents will naturally emerge. AGI itself is the strongest Agent—it will develop its own tools, call its own APIs, and complete tasks on its own.
Forcing Agents before AGI is like asking a primary school student to write a doctoral thesis. It's not a lack of effort; it's a lack of capability.
So the AI companies that are actually making money now are all adopting the human+AI model: Claude Code, Cursor, Copilot, Grok.
Not the Agent model where AI works entirely on its own.
Meta's admission that Agent progress is below expectations is essentially a calibration of the entire industry's understanding of AI's capability boundaries.
What caused the semiconductor sell-off is not the demand logic, but the narrative of unlimited expansion being punctured. $META $NVDA $MU