Mark Zuckerberg has publicly stated that AI agent development at Meta is proceeding more slowly than initial forecasts suggested. The admission comes from one of the labs that has invested most heavily in agentic systems, and it confirms what many developers are seeing in the field: getting an agent to work well on real-world tasks is harder than it looks in demos.

Why this matters for you. If you’re building workflows with agents, this statement tells you two things. First: the constraints you’re hitting—agents that get stuck, lose track, need supervision—aren’t your fault or a setup issue. They’re limitations of the technology today, even for those with the most resources. Second: agentic systems work better when you keep them focused on small, well-defined tasks with close human oversight. Don’t expect an agent to replace hours of autonomous work: use it to automate single repetitive steps, and keep everything else under your watch.

If you’re working with agents, design workflows that already account for human intervention at critical points. Architecture matters more than model power: a short chain that works beats a complex system that gets stuck.