Context compression is a huge bottleneck for long-running agents. This work finds that context compression hurts long-horizon agents at a few specific points. They propose PAIR, which replays the agent from the same state with and without a given compression, instead of

→ View original post on X — @dair_ai