Mark Zuckerberg told Meta staff this week that AI agents haven't progressed as quickly as he hoped. Meanwhile, Anthropic is pivoting Claude toward drug development and scientific research, essentially betting that the AI agency problem dissolves if you give the model a narrow enough lane. Both moves share a silent admission: general-purpose AI agents, the ones that were supposed to replace knowledge workers wholesale, are still mostly elaborate autocomplete with delusions of autonomy.
The Explainability Problem Nobody Wants to Name
A 2026 paper on arXiv by Pavel Iakovets et al. introduces PACE, a neuro-symbolic framework for counterfactual explanations, which tries to answer a deceptively simple question: if an AI made a different decision, what would have had to change? It's a framework for making machine learning legible to the humans it's supposed to serve. The reason this matters to the Zuckerberg admission is structural. Agents fail not just because they hallucinate, but because they can't model the gap between what they did and what was wanted. Counterfactual reasoning, the ability to ask "what if," is precisely what makes human judgment non-trivial to automate.
Narrow Wins, General Losses
Anthropic's Claude Science play is smart precisely because it abandons the general-agent dream in favor of constrained scientific workflows. Pulling fragments from literature, structuring hypotheses, flagging contradictions: these are tasks with ground truth. You can check them. The same logic explains why the Wiola small language model paper is more interesting than the latest GPT benchmark race. Smaller, more targeted models with legible architectures may be where the actual productivity gains land. Zuckerberg's frustration is the frustration of someone who bet on the general when the narrow keeps winning.