AI Agents Get Smarter While Using Less Power

Artificial intelligence systems that search for information online just became dramatically more efficient and capable. Researchers have developed a new approach that allows AI agents to solve complex problems while using significantly less computational power—a breakthrough that could make advanced AI more accessible and environmentally sustainable.

The key finding shows that AI agents can achieve up to 62% better performance on challenging information-seeking tasks while reducing computational costs by 10-30%. This improvement comes from a smarter way of managing how AI explores information and synthesizes answers, addressing fundamental inefficiencies in current systems.

The methodology involves a two-stage process called PARALLEL M USE. In the first stage, called Functionality-Specified Rollout, the system identifies where the AI is most uncertain during its search process and focuses computational resources on those critical moments. Instead of restarting searches from scratch each time, it reuses previously gathered information, much like a human researcher would build on earlier findings rather than repeating the same work.

In the second stage, Aggregation, the system compresses the multiple search paths into a structured report that preserves only the essential information needed to reach an answer. This is similar to how a skilled journalist might distill hours of research into a concise article, keeping the crucial facts while eliminating redundant details.

The results demonstrate consistent improvements across multiple AI models and challenging benchmarks. On the BrowseComp benchmark, which tests deep web search capabilities, the approach improved performance by 62% while requiring only 70-90% of the computational resources used by conventional methods. The system achieved near-lossless compression of reasoning trajectories, reducing context usage by up to 99% while maintaining answer quality.

This breakthrough matters because it addresses a critical bottleneck in AI development: the enormous computational costs of running advanced AI systems. As AI agents become more sophisticated, their energy consumption has grown exponentially. The new approach shows that smarter algorithms, not just more powerful hardware, can drive progress. For regular users, this could mean more capable AI assistants that are cheaper to operate and more environmentally friendly.

The technology has immediate applications in research assistance, customer service, and educational tools where AI needs to search for and synthesize information from multiple sources. It could make advanced AI capabilities available to smaller organizations and researchers who can't afford the massive computational resources currently required.

However, the approach has limitations. The research focused primarily on question-answering tasks with a limited set of tools (Search and Visit functions). Real-world applications often involve more complex interactions and broader tool sets, which could present new challenges. The method's effectiveness in scenarios requiring creative problem-solving or handling substantially larger search spaces remains an open question for future research.

The work demonstrates that fundamental improvements in AI efficiency are still possible through better algorithm design. By understanding and leveraging the unique characteristics of information-seeking tasks, researchers have shown that we can make AI both more capable and more sustainable—a crucial combination as artificial intelligence becomes increasingly integrated into our daily lives.