Agora’s Qianze Zhang Proposes Dual-Agent Architecture for Natural AI Workflow Interaction
English summary
At the second Seamate RTE monthly event, Agora Agent Platform Lead Qianze Zhang argued that current AI agents conflate communication and execution, hindering natural interaction. He proposed a dual-agent architecture where a dedicated “communication brain” handles conversation while an “execution brain” (e.g., Claude Code) performs tasks, sharing context to enable real-time user interruption and calibration. Zhang also demoed embedding Agora’s Voice AI into M5Stack hardware via vibe coding, achieving portable voice interaction and a customizable pet interface. A subsequent panel featuring guests from Google Cloud ecosystem services, AI SaaS startups, and voice products discussed monthly token usage patterns (e.g., Concept’s team of 30 consumes roughly 1 billion Google tokens at ~$400/month; Vm0.ai’s 5-person team spent ~$32k on API keys last month), tool preferences, multimodal interaction progress, the emotional and reliability value of dedicated AI hardware, bottom-up growth strategies, and the fate of SaaS. Panelist Ethan Zhang noted that workflow SaaS like Linear and Zapier are being replaced by agents, while content-input/output tools like Gamma survive, and seat-based pricing is declining in favor of usage-based models.
Chinese summary
在Seamate RTE主题月第二场活动中，Agora Agent平台负责人张千泽指出当前AI Agent常将沟通与执行混为一体，导致交互不自然，并提出双Agent架构：由独立的“沟通脑”负责对话，而“执行脑”（如Claude Code）处理任务，两者共享信息以支持用户实时打断与校准。他还展示了通过Vibe Coding将Agora语音AI嵌入M5Stack硬件，实现可随时语音对话和宠物互动界面的便携硬件。随后的圆桌讨论邀请了Google Cloud生态服务商、AI SaaS创业者和语音产品领域嘉宾，分享了月token用量数据（Concept团队30余人每月消耗约10亿Google token，成本约400美元；Vm0.ai五人团队上月API Key开销约3.2万美元）、工具偏好、多模态沟通仍需突破、硬件带来的可靠感与情绪价值、自下而上的增长策略以及SaaS的未来。嘉宾Ethan Zhang认为Linear、Zapier等工作流SaaS将被Agent取代，而Gamma等内容输入输出工具和按用量计费的模式有望存续。
Key points
Qianze Zhang proposed separating AI agents into a “communication brain” and an “execution brain” to enable continuous, interruptible collaboration.
张千泽提出将AI Agent拆分为“沟通脑”和“执行脑”，以实现可持续、可打断的自然协作。
Zhang embedded Agora Voice AI into M5Stack hardware using vibe coding, enabling hands-free voice interaction and a customizable pet display.
张千泽通过Vibe Coding将Agora语音AI集成到M5Stack硬件中，实现免提语音交互和可定制的宠物形象显示。
Panel revealed diverse token usage: Concept’s 30+ person team uses ~10B Google tokens/month for $400; Vm0.ai’s 5-person team spent ~$32k on API keys last month.
圆桌披露了不同团队的token消耗：Concept 30余人月消耗约10亿Google token，成本400美元；Vm0.ai五人团队上月API Key开销约3.2万美元。
Ethan Zhang highlighted that workflow SaaS like Linear and Zapier are being replaced by agents, while content I/O tools (e.g., Gamma) and usage-based pricing persist.
嘉宾Ethan Zhang指出Linear、Zapier等工作流SaaS正被Agent取代，而Gamma等内容输入输出工具和按用量计费模式得以存续。
Participants stressed that voice-first, multimodal interaction increases the depth of context transfer and output quality, but natural bidirectional multimodal AI is still under development.
与会者强调语音优先的多模态交互能传递更多语境、提升输出质量，但真正的双向自然多模态AI仍处于探索阶段。