For tool-heavy agents, keep only the last ~5 tool call/response pairs plus a summary

Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents (arXiv 2606.10209)·medium signalXBlueskyLinkedInCopy link

'Less Context, Better Agents' (arXiv 2606.10209) benchmarked four GPT-5 context strategies on 50 enterprise expense tasks over MCP tools and found that pruning to the last 5 tool call/response pairs and adding automated summarization hit 91.6% complete itemization and 99.64% average amount — outperforming full conversation history while slashing token cost. The lesson for builders wiring agents to verbose enterprise APIs (ERP, finance, ticketing): recent tool interactions are more decision-relevant than full history, so default to a small sliding window plus a running summary rather than appending every raw tool response.