Agent-Omit: Adaptive Context Omission for Efficient LLM Agents

Managing agent context (e.g., thought and observation) during multi-turn agent-environment interactions is an emerging strategy to improve agent efficiency. However, existing studies treat the entire interaction trajectories equally, overlooking the thought necessity and observation utility varies across turns. To this end, we first conduct quantitative investigations into how thought and observation affect agent effectiveness and efficiency. Based on our findings, we propose Agent-Omit, a unified training framework that empowers LLM agents to adaptively omit redundant thoughts and observations. Specifically, we first synthesize a small amount of cold-start data, including both single-turn and multi-turn omission scenarios, to fine-tune the agent for omission behaviors. Furthermore, we introduce an omit-aware agentic reinforcement learning approach, incorporating a dual sampling mechanism and a tailored omission reward to incentivize the agent's adaptive omission capability. Theoretically, we prove that the deviation of our omission policy is upper-bounded by KL-divergence. Experimental results on five agent benchmarks show that our constructed Agent-Omit-8B could obtain performance comparable to seven frontier LLM agent, and achieve the best effectiveness-efficiency trade-off than seven efficient LLM agents methods. Our code and data are avaliable at https://github.com/usail-hkust/Agent-Omit.

When AI assistants solve complex tasks, they often interact with tools and websites over multiple steps. During this process, they generate a lot of internal reasoning and accumulate lengthy responses from their environment — much of which becomes unnecessary as the task progresses. This redundancy wastes computational resources and slows down the AI. In this work, we first analyze which parts of an AI agent's reasoning and environmental feedback are actually useful at each step. We find that intermediate reasoning is often redundant once an initial plan is made, and early environmental responses become irrelevant by the time the agent reaches its final answer. Based on these findings, we develop Agent-Omit, a training framework that teaches AI agents to selectively skip unnecessary information during their problem-solving process. Our approach first provides the agent with examples of what to skip, then uses reinforcement learning to help it learn when skipping is safe. Experiments show that our trained agent achieves accuracy comparable to much larger AI systems while using significantly fewer computational resources.