OAgents - OPPO's open-source foundational agent framework

What are OAgents?

OAgents is an open-source foundational agent framework launched by OPPO PersonalAI Lab. Based on a standardized evaluation protocol and modular design, the framework drives research in agent frameworks. OAgents utilizes empirical system research to analyze the performance impact of different design choices for key agent components (such as planning, tool usage, and memory), and introduces a more robust evaluation protocol. The framework achieved a top average score of 73.93% on the GAIA benchmark, demonstrating excellent performance across various task levels and proving the effectiveness and robustness of its design. OAgents supports the integration of multiple agent components, providing a solid foundation for future research.

Main functions of OAgents

• Multimodal tool integrationOAgents integrates multimodal tools for processing text, voice, images, and video, allowing direct interaction with multimodal inputs and enhancing the ability to acquire and interpret factual information in complex real-world scenarios.
• Optimized Search AgentBased on optimized multi-source retrieval, query optimization, and a simplified browsing architecture, OAgents can effectively perform web searches, expand knowledge boundaries, and provide more accurate information support for complex tasks.
• Dynamic Programming and Task DecompositionBy using dynamic programming, complex tasks can be broken down into executable subtasks, and the plan can be dynamically adjusted based on real-time observations, thereby improving the management capabilities and reasoning efficiency of complex tasks.
• Memory Enhancement Knowledge SystemOAgents constructs hierarchical memory modules, including current memory, memory summary, vectorized retrieval, and long-term memory, to enhance the agent's cognitive abilities and help the agent perceive, reason, and make decisions more effectively in complex environments.
• Expansion strategy during testingDuring the testing phase, OAgents employs extended strategies, such as diversity enhancement, optimization, and reward modeling, to dynamically adjust the decision-making process, thereby improving the agent's adaptability and exploratory capabilities and enhancing overall performance.

OAgents' technical principles

• Multimodal tool principleOAgents' multimodal tools convert non-text content into text descriptions and simultaneously perform cross-modal semantic parsing, enabling direct interaction and understanding of multimodal inputs. The formula is expressed as: Response = A(xtext,Timage(I),Tvideo(V)), where A is the Agent function, xtext is the text input, and Timage and Tvideo are the tool functions for images and videos, respectively.
• Search AgentprincipleOAgents' search agent framework integrates commercial APIs and archive systems to achieve multi-source retrieval; it optimizes queries through a closed loop based on semantic calibration and morphological expansion; and it simplifies queries into three atomic functions: search, access, and read, reducing complexity.
• Dynamic Programming PrincipleOAgents' dynamic programming module generates high-level plans, breaking down tasks into executable steps and periodically revising the plans based on new observations during execution to adapt to dynamic environments. OAgents employs hierarchical task decomposition, constructs dependency graphs, and dynamically schedules executable subtasks.
• Memory enhancement principleOAgents' memory module stores short-term information based on current memory, summarizes and extracts high-value knowledge using memory, quickly retrieves relevant historical memories based on vectorized retrieval, and integrates historical insights using long-term memory to optimize task execution.
• Expansion principle during testingOAgents' test-time extension module uses a hybrid sampling strategy to enhance diversity, optimizes decision paths based on process-based reward functions, and employs a real-time reflection mechanism for adaptive problem solving.

OAgents project address

• GitHub repositoryhttps://github.com/OPPO-PersonalAI/OAgents
• arXiv technical paper: https://arxiv.org/pdf/2506.15741

Application scenarios of OAgents

• Intelligent Customer ServiceWe can answer customer questions quickly and accurately, provide personalized solutions, handle complex customer inquiries, and improve customer satisfaction.
• Educational guidanceIt provides students with personalized learning plans, dynamically adjusts teaching content based on learning progress and feedback, handles various learning materials, and offers a rich learning experience.
• Medical consultationIt assists doctors in analyzing medical records, making diagnostic recommendations and developing treatment plans, accessing the latest medical research and clinical guidelines, and providing data-driven decision support.
• Smart office assistantIt assists users in scheduling, writing reports, and organizing meeting minutes, remembering user preferences and habits, and providing personalized office support.
• Smart Home ControlIt integrates multiple smart home devices, controls devices based on voice or text commands, realizes automated scenarios, and provides a natural interactive experience.