Story

In brief

We ran a single agent session through all 89 sequential tasks of Terminal Bench 2.0 or over 80 million tokens, with no measurable accuracy loss versus running each task in its own fresh session. We didn't use compacti...