OpenAI's agents hit a trust bottleneck

This week's frontier AI story is about what happens when agents run for hours. OpenAI says its internal research agents now complete 3.1 agent-workdays of effort per human workday. 1

OpenAI also reports that more than half of successful agent tasks lasting 4 to 8 hours still needed at least one human intervention. 1

OpenAI Chief Scientist Jakub Pachocki has warned that recursive self-improvement could arrive while alignment and monitoring lag. 2 OpenAI's "An Alien Mind" explains why chain-of-thought monitoring becomes less dependable as models use tools and manipulate their own reasoning. 3

Agent use is moving forward, but trust and monitoring remain the bottleneck. 123