Meta's AI agent push lagging, Zuckerberg says; Watermelon

Meta acknowledged at an internal town hall on Thursday that the company’s work on AI agents has not progressed as quickly as executives expected, and that the reorganization around agentic systems "haven't come to fruition yet." The admission came as CEO Mark Zuckerberg said the "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected."

What did Zuckerberg say about agent progress?

Meta's agent roadmap has underperformed recent expectations, Zuckerberg said, and executives misjudged timing; he expects more tangible results within the next three to six months. The company reorganized around agents earlier this year, moved about 7,000 employees into AI teams after a roughly ten percent global workforce reduction in May, and announced plans to spend up to $145 billion on AI infrastructure this year.

Zuckerberg described planning that began in January and February, when senior leaders worried they were not moving fast enough and were "super optimistic" about third-party tools. He framed the last four months as a period in which agentic development did not accelerate as anticipated.

How has Meta rearranged around agents and models?

Meta shifted staff and budgets to prioritize AI: the company laid off roughly ten percent of its global workforce in May and reassigned about 7,000 employees to AI roles, while preparing large infrastructure investment. It rebranded its AI division under Alexandr Wang as Meta Superintelligence Labs and paid aggressively to recruit talent, offering nine-figure packages to some hires.

Product work continues in public and internally. In April Meta released Muse Spark, the first model in a new lineup; it produced solid benchmark scores but did not match OpenAI or Anthropic, according to the company’s internal discussion. Alexandr Wang told staff that an internal successor, code-named Watermelon, is in training and uses "an order of magnitude more compute than Avocado," the internal name for Muse Spark.

What did Meta's AI chief say about model progress?

Alexandr Wang pushed back on the notion that Meta's AI effort is falling behind, saying internally that Watermelon has caught up with OpenAI's top model GPT-5.5 on unspecified benchmarks and that it is currently in training. Wang told staff Watermelon uses far more compute than Avocado, and he signaled on X that a Muse Spark update with major improvements to coding and agentic capabilities is coming soon, followed by a coding model he said will be on par with Anthropic's Claude Opus "pretty soon."

Wang framed Zuckerberg's remarks as referring to the broader industry's pace rather than Meta's internal progress, and said users will like what the team has been "cooking."

What remaining operational issues did leadership address?

Meta paused an internal mouse-tracking program that recorded employee mouse movements and digital activity for AI training, and CTO Andrew Bosworth said an internal review found no employee data entered AI training sets. Bosworth acknowledged that when the software was first installed on U.S. employees' machines in April, he had told them there was no way to opt out. If the program restarts after the review, Bosworth said it would run on an opt-in basis. He added, "For people who are comfortable, that's great, they can contribute to this kind of great human survey," and reassured others that "it is not an issue."

Why it matters

Meta has concentrated people and capital behind agentic systems: roughly 7,000 staff redeployed, a May workforce cut of about ten percent, and a declared infrastructure budget of up to $145 billion this year. If agentic capabilities lag while such resources are consumed, Meta risks a period in which heavy spending and structural disruption produce fewer short-term returns. The split public messages from Zuckerberg and Alexandr Wang highlight internal tension on progress and timing.

What to watch

Track model releases and concrete benchmark disclosures: Meta says a Muse Spark update and a new coding model are coming, and Wang claims Watermelon is training. More measurable signals will be publication of benchmarking numbers, delivery dates for Watermelon and the coding model, and whether the mouse-tracking program restarts on an opt-in basis.