In 2025, expectations for the rise of “AI agents” were high, as indicated by Nvidia CEO Jensen Huang. Notable developers in the AI sector, including OpenAI, Google, and Alibaba, have introduced refined AI models tailored for specific tasks like web search and report generation. However, a significant barrier to the successful implementation of these AI agents is their ability to remain focused during lengthy tasks, especially as benchmark tests reveal increased failure rates with the number of steps taken and prolonged time frames.
A new academic framework, EAGLET, aims to enhance the long-term task performance of language model-based agents without requiring manual data labeling or retraining. Developed by researchers from Tsinghua University, Peking University, DeepLang AI, and the University of Illinois Urbana-Champaign, EAGLET introduces a “global planner” designed to integrate seamlessly into existing workflows to minimize errors and increase efficiency.
EAGLET operates by creating a high-level plan from task instructions but does not interfere during task execution. This preparatory guidance helps reduce planning errors and increases success rates for the agents. The framework is structured with a two-stage training pipeline that generates synthetic plans without the need for human-generated content.
One notable component is the Executor Capability Gain Reward (ECGR), which assesses generated plans based on their effectiveness for varying capabilities among agents while promoting shorter, more efficient task pathways. EAGLET has shown compatibility with existing AI models, achieving performance improvements across several foundational models, including GPT-4.1 and GPT-5.
In evaluations against three benchmarks that simulate real-world tasks, agents utilizing EAGLET outperformed those using traditional methods, demonstrating significant enhancements in task completion rates and reduced step count. Despite these advancements, there are questions regarding the accessibility of EAGLET for enterprise deployment, as no public code has been released yet. Further inquiries have been raised about the ease of integration with existing enterprise frameworks and the potential for minimal model scalability. Additionally, there remains uncertainty about the most effective deployment strategy for EAGLET, whether for real-time execution or pre-generated planning.
Source: https://venturebeat.com/ai/eaglet-boosts-ai-agent-performance-on-longer-horizon-tasks-by-generating