Even Zuckerberg Admits It: AI 'Agents' Are Harder to Build Than Promised
In a leaked internal meeting, Mark Zuckerberg admitted Meta's push to build AI 'agents' has stalled — despite a $145 billion bet and a painful reorganization. A useful reality check on one of the year's biggest buzzwords.
You’ve probably been hearing the word “agents” a lot this year. It’s the industry’s term for AI that doesn’t just answer questions but actually does things for you — books the appointment, sorts the inbox, finishes a multi-step job on its own. So it’s worth noting when the CEO who bet his entire company on that idea quietly admits it isn’t going to plan.
In a leaked internal town hall (Reuters got the audio), Mark Zuckerberg told Meta staff that “the trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected.” This from a company that reorganized itself around AI — laying off about 10% of its workforce in May and moving roughly 7,000 people onto AI teams. Zuckerberg acknowledged that leadership had been “super optimistic” when they drew up those plans early in the year.
Why are agents so hard? The honest answer is that doing a chain of tasks reliably is a different beast from answering a single question. Every step is a chance to go slightly wrong, and small errors compound — an agent that’s 95% reliable per step can still fail most ten-step jobs. Agents also stumble over vague instructions and unexpected situations in ways that make them risky to trust unsupervised with anything important. Even the best agent systems from any lab in 2026 still need a human watching for consequential work. It’s a bit like self-driving cars: impressive demos came fast, but the last stretch to “trust it completely” is the hard, slow part.
A fair bit of perspective: Meta is spending up to $145 billion on AI infrastructure this year, so a four-month plateau is a timing wobble, not a change of heart. Zuckerberg still expects real progress in three to six months. But the gap between this year’s confident promises and this quarter’s reality is exactly the kind of thing worth remembering the next time a product claims its agent will run your life.
What this means for you: If you’ve felt behind because you’re not yet handing tasks to autonomous AI — relax. The company spending more than almost anyone is finding it harder than advertised. For now, AI is genuinely useful as a capable assistant that you supervise, and shaky as an unsupervised employee. If your business is scoping “let AI handle it” workflows, build in extra time, keep a human in the loop for anything that matters, and treat bold agent timelines as marketing until proven otherwise.
Sources
MiniMax M3: A Top-Tier AI Model You Can Download and Run Yourself
A Chinese lab released M3, a free-to-download AI model that matches the big closed systems on coding tasks, reads a million words at once, and handles images and video. A look at why 'open-weight' models keep closing the gap — and the fine print worth reading.
