Andrej Karpathy told No Priors that the core bottleneck in software development is no longer a human's typing speed or ability to write code, but their capacity to effectively orchestrate and delegate to AI agents.
Karpathy revealed he hasn't manually typed a line of code since December, with his primary task now being the expression of intent to AI agents, which he described as a form of 'manifesting'.
According to Karpathy, modern engineering mastery is defined by managing multiple persistent AI agents, with a key performance metric being the effective token throughput of your orchestration system.
The show noted Peter Steinberg's setup, where dozens of parallel agents run autonomously for around 20 minutes each to complete complex coding tasks across multiple repositories.
Karpathy argued that the most important attributes in an AI agent are now personality, memory, and feedback loops, as these foster a collaborative teammate dynamic rather than a simple tool.
Karpathy built a personal agent named Dobby that runs his smart home, autonomously discovers network devices, and operates persistently in the background as part of a 'loopy' era of agent design.
No Priors highlighted a psychological pressure emerging among developers, where idle agents or an underutilized AI subscription are seen as a wasted competitive edge, akin to the historical anxiety over idle GPUs.