arXiv cs.AIWednesday · May 27, 2026FREE

AGORA: Adapter-Grounded Observation-Action Retention for Inference-Free Prompt Compression in LLM Agents

AGORA (Adapter-Grounded Observation-Action Retention) is a prompt compression technique designed specifically for LLM agents, addressing the failure of general token-level extractive compressors in agentic settings. The paper, published on arXiv on May 27, 2026, reports that across 17 (environment, backbone, method) cells spanning two independent token-level method families, every cell collapsed to a mean reward of 75% of uncompressed performance in 8 of 9 cells (with the lone exception at 73%). A four-way component ablation revealed that the structural floor—the retention of observation-action pairs—is the dominant quality lever, while the learned scorer provides 1.0-11.5x adaptive end-to-end compression from a single fixed keep ratio. AGORA requires no inference-time overhead, making it suitable for real-time agent applications. The method is evaluated on multiple agent benchmarks, demonstrating that it preserves task performance while significantly reducing prompt length.