Effect AI

First large-scale study of AI agents in production just dropped: 306 practitioners, 20 case studies, 26 domains. (Pan et al, Berkeley + collaborators, arXiv:2512.04123) The findings that should reframe every "agentic AI" pitch you sit through this quarter: • 68% of production agents execute ≤10 steps before human intervention • 70% use prompting off-the-shelf models, NOT fine-tuning • 74% rely primarily on human evaluation • 66% are deployed in latency-tolerant workflows The top challenge isn't model intelligence. It's reliability. The vendors selling "fully autonomous, fine-tuned, instant-response" agents are pitching the opposite of what's working in production. The winners are simple, controllable, human-in-the-loop systems with measured handoffs.