Topic

AWorld Agent Framework: From Single Model to Collective Intelligence

AWorld is a next-generation artificial intelligence framework developed by InclusionAI, designed specifically for large-scale agent self-improvement. Through self-awareness and collaborative learning, the framework enables AI agents to continuously evolve, addressing the core challenge in the AI field of "discovering their own limitations." The core innovation of AWorld lies in its three-pillar architecture: a plug-and-play protocol for multi-agent systems (MAS), intelligence generation capabilities beyond a single model, and cloud-native scalability. The framework has already demonstrated outstanding performance in practical applications, including solving 5 out of 6 problems in the 2025 International Mathematical Olympiad, and achieving Rank 1st on benchmarks such as GAIA, OSWorld, and VisualWebArena. In terms of technical architecture, AWorld provides a comprehensive component system, including agent management, runner system, task framework, group coordination, sandbox environment, tool integration, context management, memory system, and tracing framework. These components together support the full range of needs, from simple agent construction to complex multi-agent collaboration. Outline: 1. Why build agents 2. Definition of an agent 3. How agents have developed 4. Practice at Ant Group 5. Achievements and awards of agents 6. Specific introduction of framework algorithms 7. Future outlook