According to Lance Martin, an engineer on Anthropic's agent infrastructure team, AI agents can now run unattended for over twelve hours, a leap from about ten minutes of autonomous work three years ago. Martin detailed the architectural breakthroughs on the AI Engineer podcast, emphasizing that the shift is due to a fundamental redesign of infrastructure around models, not just better models. In 2024, Claude 3 Opus could sustain roughly ten to twenty minutes of autonomous work, limiting products to chat and autocomplete. By 2025, models crossed the one-hour threshold, enabling synchronous coding agents like Claude Code. Starting in April 2026, Anthropic's "Mythos class" models pushed past twelve hours, shifting the product surface from synchronous to fully asynchronous. Martin noted that the jump demands an architecture where the system can survive crashes, verify work without hallucinating, and correct persistent errors. Anthropic's first attempt to build Managed Agents placed orchestration logic and the sandbox inside the same container, but this design failed because when the container died, the entire session was lost, and credentials inside the container gave the model unsupervised access to secrets, a serious security concern.