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

In three linesStudy on context optimization for autonomous LLM agents in enterprise workflows. Testing 4 GPT-5 configurations on 50 expense itemization tasks (Microsoft Dynamics 365). Pruning context to last 5 tool calls + summarization achieves 91.6% completion with 553k tokens (vs 1.48M full context), reducing runtime from 14.56h to 5.79h.

Summary generated by Claude — human-verified