I’m kind of curious about this:
“ Long-horizon tasks: We are increasingly solving tasks that take hours of human equivalent time. Agents are now able to maintain capability even as the context window fills with thousands of tool calls. The human-equivalent time-horizon continues to grow”
Because the most efficient humans don’t do this, they have short cycle times. This “agents can operate without intervention” always felt like an anti-pattern to me.
