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

Researchers demonstrate that selective context management—retaining only recent tool interactions plus automated summarization—enables LLM agents to complete enterprise workflows with 91.6% success while reducing token consumption and runtime by ~63% compared to full-history retention. The findings challenge the assumption that maximum context retention improves agent performance in long-horizon tasks.

The research addresses a critical inefficiency in deployed LLM agents: verbose tool responses from enterprise systems create exponential growth in context windows, driving up costs and latency while paradoxically degrading performance. Using Microsoft Dynamics 365 expense itemization as a testbed, the study reveals that full-context retention achieves only 71% task completion despite consuming 1.48M tokens per benchmark run. The counterintuitive finding—that aggressive pruning combined with summarization outperforms exhaustive context retention—stems from how language models process information. When context windows overflow, models struggle with stale-state errors where earlier information contradicts recent tool responses, and attention mechanisms dilute focus across irrelevant historical exchanges. By retaining only the last five tool interactions plus compact summaries, the approach maintains task-relevant information while eliminating noise that confuses the model's decision-making. This 91.6% completion rate with 553K tokens represents a meaningful efficiency gain for production systems handling thousands of transactions daily. The research extends beyond expense management; it establishes a replicable methodology for context engineering that applies to any enterprise workflow involving stateful tool interactions. Cross-model validation with Claude Sonnet confirms findings aren't GPT-specific. Organizations currently deploying agents with retention-heavy strategies face immediate optimization opportunities. The work also signals that future agent architectures should embed selective context management rather than treating it as an afterthought, reshaping how LLM frameworks approach memory and state management.

• →Selective context retention plus summarization achieves 91.6% task completion versus 71% with full-history retention, while reducing tokens by 63%
• →Context overflow causes stale-state errors and attention dilution that paradoxically harm agent performance despite providing more information
• →The last-5-interactions pruning window emerges as an optimal balance, improving completion to 79% while halving computational costs
• →Methodology applies broadly to enterprise workflows with stateful tool interactions, not limited to expense processing
• →Cross-model validation with Claude Sonnet 4.5 confirms findings generalize beyond GPT architectures