LLM Agent State Compression Impacts Constraint Preservation in Relays

In long-running Large Language Model (LLM) agent systems, intermediate states often grow large, necessitating compression before being passed to subsequent decision-making steps. This compression creates an information bottleneck where critical details can be lost, potentially breaking strict numeric or categorical constraints. This study investigates the impact of different state compression techniques on constraint preservation within a two-agent LLM relay system, using a closed-world travel-planning scenario. The experiment involved a "Researcher" agent auditing hotel and flight inventory and a "Booker" agent selecting a pair based solely on the goal and the compressed hand-off payload. Four hand-off conditions were compared: no compression, narrative summarization, schema-constrained JSON extraction, and embedding-based pruning. The results showed that the hand-off representation critically affects downstream feasibility. JSON extraction achieved the highest feasibility accuracy at 0.96, while narrative summarization, despite being the most compact, degraded feasibility to 0.48. Embedding-based pruning performed well, matching the uncompressed control at 0.88 without an additional generative compression call. These findings underscore that structured, auditable hand-off representations are superior to mere brevity for preserving constraints.