In September 2013, Elon Musk said Tesla should be able to drive “90 percent of miles” autonomously within three years, meaning by 2016. At the time, the promise felt realistic because the early pieces were already working: lane keeping, highway assistance, sensor fusion, and assisted control had made autonomy look deceptively close.
But the next decade showed that achieving higher levels of reliability was very hard. As Andrej Karpathy recently said on the Dwarkesh Podcast that autonomy is a “march of nines.” A solution that works 90% of the time is only the first nine; every additional nine takes roughly the same amount of work.
AI agents are at the same stage now
The reliability problem is now moving to agents and multi-agent workloads, because this is where AI becomes economically useful. It is one thing for a model to answer a question. It is another for a customer service agent to spin up sub-agents, investigate a complaint, retrieve customer data from a CRM, negotiate a retention offer, process a payment, and update the customer’s plan. At every step, these sub-agents have to choose the right tool, pass the right parameters, interpret the result correctly, and hand it off without losing or inventing context.
Models are improving every few months, but even frontier agents are only around 82% correct on popular tool-use benchmarks, and when stitched together into multi-agent workflows, failure rates can creep north of 80% [MCP-Atlas, 82.2%; Cemri et al., 41.0-86.7% MAS failure rates]. Despite that reliability gap, agents are too useful to ignore. Individuals and enterprises want agents operating inside their day to day workflows, and that demand is only going to grow. To give a sense of scale, agentic work will soon exceed pure reasoning work; by 2030, as many as one in three tokens may be used for tool-calling tasks ().
