Less Context, Better Agents: Study Shows AI Tools Work Smarter by Reducing Clutter

Researchers from Microsoft and other institutions published a study on arXiv showing AI agents perform better when they use less context. The team tested four versions of GPT-5 on expense-processing tasks in Microsoft Dynamics 365, using the Model Context Protocol (MCP) tools. Some versions kept full conversation histories, while others focused only on the most recent tool responses or a pruned set of the last five tool call/response pairs. The agents that ignored older, less relevant information made fewer mistakes, reduced context overflow and stale-state errors, and processed expenses more efficiently at lower inference cost.

This discovery matters because it could make AI assistants in business tools like accounting software or customer service platforms more reliable and cost-effective. Right now, these AI agents often get confused by too much information, especially when verbose tool responses from enterprise systems cause context overflow. By focusing only on what's immediately useful, they can avoid errors and work faster - like how you'd focus on the current task instead of remembering every conversation you've ever had.

If you use AI-powered business tools, this research suggests a simple way to test if your AI assistant is working optimally: Check if it's keeping full conversation histories. If it is, try asking the support team if they can configure it to focus only on recent, relevant information. For Microsoft Dynamics 365 users, look for 'Model Context Protocol' settings in your admin panel.