PAIR framework targets context compression failures in long-horizon agents

A [research paper](https://arxiv.org/abs/2609.36526) highlighted by DAIR.AI investigates performance bottlenecks in long-horizon agents, finding that major drops in success rates originate from a small number of specific compression events rather than gradual degradation. Harmful compressions typically discard unresolved task criteria or reduce technical specifications into vague text, forcing agents to repeat API authentication and documentation lookups.

To resolve these failures, the researchers created PAIR, which replays the agent from the same execution state with and without a given compression event instead of comparing noisy full runs. PAIR diagnoses the specific missing context and rewrites targeted sections of the compression prompt while keeping the agent model, compressor, and tool interfaces fixed. Tested on AppWorld, OfficeBench, and tau-Bench Retail, the method delivered the most consistent completion rates among compressed approaches, approaching uncompressed execution baselines.