# Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents

**Authors:** Abhilasha Lodha, Mahsa Pahlavikhah Varnosfaderani, Abir Chakraborty, Abhinav Mithal  
**arXiv:** [2606.10209](https://arxiv.org/abs/2606.10209)  
**Date:** 2026-06-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2606.10209v1/

## Key points

- Authors address challenges faced by large language models deployed as autonomous agents in enterprise workflows
- Focus on verbose tool responses leading to context overflow, stale-state errors, and high inference cost in automated expense itemization within Microsoft Dynamics 365 Finance and Operations using Model Context Protocol tools
- Evaluation of four different configurations of GPT-5 on a 50-task hotel expense benchmark
- No user model achieves only 8.0% complete itemization
- Retaining full conversation history improves completion to 71.0%, but at a significant cost
- Pruning context to last 5 tool call/response pairs enhances completion to 79.0%
- Adding automated summarization to pruning recent interactions achieves a completion rate of 91.6% for itemization and an average amount itemized of 99.64%
- Selective retention of recent tool interactions combined with compact summarization significantly enhances reliability and efficiency compared to retaining full-history context

## Summary

In their study titled "Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents," authors Abhilasha Lodha, Mahsa Pahlavikhah Varnosfaderani, Abir Chakraborty, and Abhinav Mithal address the challenges faced by large language models deployed as autonomous agents in enterprise workflows. They focus on the issue of verbose tool responses from enterprise systems leading to context overflow, stale-state errors, and high inference cost specifically in automated expense itemization within Microsoft Dynamics 365 Finance and Operations using Model Context Protocol tools. The research evaluates four different configurations of GPT-5 on a 50-task hotel expense benchmark. These configurations include no user model, full conversation history retention, context pruned to the last 5 tool call/response pairs, and pruning with automated summarization. The results are based on averaging across 5 independent runs while keeping the user model constant for comparison in context engineering. The baseline scenario without a user model achieves only 8.0% complete itemization. Retaining full conversation history improves completion to 71.0%, but at a significant cost of consuming 1,480,996 tokens and taking 14.56 hours per benchmark. Pruning the context to only the last 5 tool calls enhances completion to 79.0%, while reducing token usage to 535,274 and runtime to 5.39 hours. The most successful configuration involves adding automated summarization to pruning recent interactions which achieves a completion rate of 91.6% for itemization and an average amount itemized of 99.64%. This is accomplished with a token use of 553,374 and a runtime of 5.79 hours. Additionally, the study reports confidence intervals, effect-size analysis, sensitivity over pruning and summary windows, failure analysis, results across five expense types grouped into three categories, and cross-model evidence with Claude Sonnet 4.5. Overall, the findings demonstrate that selective retention of recent tool interactions combined with compact summarization can significantly enhance both reliability and efficiency compared to retaining full-history context in this class of enterprise tool-use workflows.