• The most important current agent weakness, in Mantinband's view, is handling huge amounts of context. 2:58
  • Even a 1 million token context window is not enough for a reliable grasp of a large production codebase. 3:23
  • Because an agent may miss whole parts of a codebase, Mantinband says developers cannot blindly trust it to produce the correct implementation for a plan. 3:23
  • He expects context handling to improve, but places it in the "humans have some time" category because larger context windows are expensive in memory and GPU usage. 3:53
  • He calls out codebases of roughly 300,000+ lines as a scale where context windows and compression remain a serious bottleneck. 4:28
  • His rough estimate is that this context advantage may remain meaningful for at least another one or two years, though he acknowledges hardware and tooling could change that. 4:57
  • He sees agent vision as a weaker source of human advantage because multimodal capabilities are improving quickly. 4:57 5:23
  • He says agents remain bad at coming up with novel solutions, especially in complex domains where a simple original idea is needed. 5:45
  • The durable human advantage is not typing code but understanding the large product picture and forming novel architectural or domain ideas. 7:25