On October 23, the iRTE2026 Real-Time Intelligence Conference was held in Beijing. In the main forum, Agora founder and CEO Zhao Bin delivered a speech titled "Conversational AGI: The Next Chapter of Real-Time Interaction," reviewing the key changes in real-time interaction, conversational AI, and underlying infrastructure over the past year, and proposing a "five-level conversational capability" framework that outlines the evolution of conversational AI from basic question answering toward deep understanding and relational communication.
Zhao Bin believes that as AI moves from answering questions to continuous conversation, understanding context, and executing tasks, the objects of real-time interaction are expanding from person-to-person to person-to-Agent. The role of real-time interaction technology is also extending accordingly: it must not only make information transmission more stable and smoother, but also provide real-time support for AI's perception, understanding, expression, and action.
From "One Question, One Answer" to Deep Relational Communication: The Five-Level Evolution of Conversational AGI
In Zhao Bin's view, the progress of conversational AI should not be measured only by response speed and language naturalness. More important is whether it can understand the intentions, circumstances, and relationships of its conversation partners, whether it can express itself appropriately in complex scenarios, and whether it can gradually form long-term, deep understanding.
Based on this judgment, Zhao Bin proposed the "five-level conversational capability" framework in his speech, dividing conversational capability into five levels:
L1, expression based on one's own perspective: mainly self-expression and passive continuation, achieving "one question, one answer."
L2: expression based on interaction rules: moving from passive response to active expression, making conversations more natural and coherent.
L3: communication based on individual cognition: understanding an individual's identity, intentions, and circumstances, achieving communication tailored to each person.
L4: communication based on group cognition: understanding social relationships and role differences, adapting to complex multi-party communication scenarios.
L5: communication based on deep relationships: understanding long-term context and relationship evolution, forming deeper understanding and tacit thinking, moving communication from understanding the moment to understanding a person.
What these five levels demonstrate is not just the step-by-step enhancement of conversational technology, but a change in the relationship between AI and humans: from completing a round of Q&A, to understanding a person, to understanding relationships between people, and ultimately forming deep communication that transcends time and context.
Zhao Bin pointed out that conversational AI is evolving from early passive, turn-based interaction toward more proactive and natural full-duplex conversation, and is beginning to attempt to understand people's personality, circumstances, emotions, and relationships. Looking ahead, AI also needs to break through role division and communication order in complex social scenarios, and further understand an individual's long-accumulated context. This is also an important development direction for conversational AGI: AI should not only understand a single sentence, but also combine complete context to understand a person's true intentions and respond appropriately.
Real-Time AI Infra: Key Infrastructure Supporting Conversational AGI
The continuous evolution of conversational capability also places higher demands on underlying technology. Zhao Bin mentioned in his speech that over the past year, the deep adoption of WebRTC technology by OpenAI and other companies has driven broader attention to and application of real-time interaction technology in the field of conversational AI. As AI moves from single-turn Q&A to continuous interaction, real-time interaction infrastructure is becoming an important technological bridge connecting models, agent runtime frameworks, and users.
Compared with traditional AI applications, conversational scenarios impose stricter requirements on latency, stability, and two-way interaction. AI needs to receive user input while understanding context and organizing responses, and when necessary respond to interruptions at any time, continuously call tools, and advance tasks. Any noticeable delay in any link may disrupt the naturalness of the conversation, or even affect whether a task can be completed.
To support these needs, Zhao Bin introduced infrastructure directions for real-time intelligent interaction: through a global real-time transmission network, distributed cloud-edge-device joint inference, real-time orchestration, and full-duplex interaction capabilities, models and Agents can collaborate continuously in real-world scenarios. Agora's Conversational AI Engine also takes full-duplex interaction as its core, connecting models, Agent runtime frameworks, and users, supporting stable, synchronized real-time conversation experiences.
At the same time, Agora continues to improve conversational AI development tools and development kits for developers, combining SDKs, Skills, CLI tools, Recipes, and other capabilities to help developers more efficiently build Agents with complete logic and execution capabilities.
From customer service and education to AI companionship and smart wearables, Zhao Bin believes conversational AI is entering more real-world scenarios. Whether AI can truly create value depends not only on whether the model "can speak," but also on whether it can understand needs, communicate naturally, and consistently complete tasks. Real-time interaction technology will play an increasingly important role in this process.
Next, Andrey Norkin, Chief Architect at NVIDIA and Co-Chair of the AOMedia Video Codec Working Group, gave a presentation on AV2 video codec technology.
Roundtable Discussion: Will Personal Agents Become the Next-Generation Personal Entry Point?
In the first roundtable moderated by Agora co-founder Tony Wang, interactive media product manager and tech blogger Eugene Wei, tech self-media personality and host of the podcast "Luan Fan Shu" Pan Lu, and AIoT hardware entrepreneur and host of the edge-AI self-media channel Hua Jianxin discussed the topic "Personal Agents: From New Tools to New Entry Points?"
From phones, glasses, and earphones to more wearable devices, personal agents are once again prompting the industry to think about the next-generation human-computer interaction entry point. Pan Lu believes personal agents may bring a paradigm shift: the value of AI products is shifting from competing for user attention to helping users complete tasks more efficiently. Future personal agents should not only respond to explicit commands, but also proactively take on tasks, helping users handle matters that have clear goals yet are not worth investing large amounts of time in.
Regarding next-generation personal hardware, Hua Jianxin proposed that the future entry point may not be monopolized by any single type of device—glasses, earphones, and other hardware capable of sensing the environment and executing tasks all have opportunities. He believes personal agents not only need to understand users, but also need to continuously perceive the real world and take action through physical devices, which brings new development space for hardware of various forms.
Eugene Wei, meanwhile, steered the discussion toward the value of "context": whether a personal agent can truly understand a user's interests, communication records, and social relationships will affect whether it can become a valuable personal entry point. In evaluating whether a new entry point is viable, he also suggested paying attention to how often users interact with the Agent each day, and whether users reach out to the Agent more often, or the Agent reaches out to users more often. The roundtable then further discussed the product value, hardware forms, and future measurement standards of personal agents.
Roundtable Discussion: From "Being Able to Chat" to "Getting Things Done"—How Does Conversational AI Create Real Value?
In the second roundtable moderated by Agora Chief Business Officer Cui Bo, Xieyue Intelligence founder Chen Wei, Guanke Technology founder Geng Xianglei, Rokid Leqi co-founder and head of ecosystem Cai Guoxiang, and Titan Qiming CEO Wu Kan explored conversational AI—driving next-generation human-Agent interaction—from different scenarios including embodied intelligence, smart wearables, enterprise marketing, and AI recruiting.
Chen Wei believes the key to embodied intelligence is truly translating human intent into robot actions. Robots must not only understand voice commands, but also fuse multimodal information such as vision, speech, and their own state to complete perception, planning, and decision-making in the physical environment, and continuously adjust task execution based on environmental changes after acting. Rather than completing pre-set demonstration tasks, what matters more for robots is continuously understanding people's real needs in open environments, keeping actions always aligned with human intent. Xieyue Intelligence is developing a Duplex Reasoning, an entirely new paradigm of embodied foundation model, aimed at unifying robot interaction and action.
Cai Guoxiang, speaking from the practice of AI glasses, emphasized that the conversational experience is the core competitiveness of AI wearable devices. He stated that no matter how powerful the hardware and the underlying AI capabilities are, if the conversation is not natural enough, the user experience will be directly affected. In the future, AI wearable devices need low-latency, stable real-time interaction, along with understanding of semantics and long-term context, so that users gradually no longer feel they are talking with AI. He also shared the practice of Rokid Leqi's AI assistant using Agora's RTC link services, and expressed optimism that AI glasses will become an important personal entry point in the AI era.
In enterprise application scenarios, Geng Xianglei pointed out that for AI voice agents to achieve reliable business delivery, they must solve problems of latency, interruption timing, and humanization—latency in particular can even directly affect task outcomes. For To B marketing, AI also needs to be combined with industry knowledge systems and human collaboration to better guarantee business results. Wu Kan, drawing on AI recruiting practice, said conversational AI can help recruiting businesses reach job seekers at scale and improve service efficiency, rather than simply replacing humans. To form a closed business loop, it is also necessary to accumulate industry data, organize the business around the job seeker as the core entity, help the Agent build long-term memory, and support AI in continuously evaluating and optimizing service quality around actual business events such as interviews and onboarding.
From the sharing by the four guests, it can be seen that although different industries have varying requirements for conversational AI, they all point in a common direction: AI must not only be able to communicate, but also understand real scenarios, integrate into existing business processes, and reliably complete tasks. As real-time interaction, multimodal perception, and Agent execution capabilities continue to develop, conversational AI is expected to evolve from an interaction tool into an important force connecting the digital and physical worlds and driving business operations.
From real-time interaction to real-time intelligence, opening the next chapter of conversational AGI
From the five-level conversational capability framework proposed by Zhao Bin, to the two roundtable discussions on personal agents and real-world business deployment, the iRTE2026 main forum presented a clear industry trend: competition in conversational AI is extending from foundational model capabilities to context understanding, real-time interaction, complex relationship cognition, and task execution.
When AI is no longer limited to passively answering questions, but can continuously understand people, actively participate in conversations, and complete tasks in real scenarios, real-time interaction technology will also take on a deeper mission.
Facing the next stage of conversational AGI, Agora will continue to explore the integration of real-time interaction and generative AI together with developers and industry partners, promoting more natural and efficient real-time interaction experiences between people and between people and Agents.
This article was provided by Agora and republished by QbitAI with authorization; the views belong to the original author.