“小狗，去找你的好朋友瓦力，打个招呼。”收到指令，机器狗立马朝瓦力走去，主动打起了招呼。再来一句，“给瓦力跳个舞吧。”好家伙，它还真跳上了。只见机器狗在展台上扭动身体，一段颇有喜感的舞蹈说来就来。找朋友、打招呼、跳舞，这只机器狗，是真听劝。10月20日上午，2026物理AI与世界模型大会在南京开幕。图灵奖得主通过视频分享前沿判断，院士与产业界嘉宾齐聚现场，一批创新成果也在展区集中亮相。在爱思科技的展区，过去要让机器狗完成这样的动作，往往离不开人工遥控。现在，用户只需说出想要的结果，剩下的工作就交给AI。
"Doggy, go find your good friend Wall-E and say hello." Receiving the command, the robot dog immediately walked toward Wall-E and greeted it. Another command: "Dance for Wall-E." Sure enough, it started dancing. The robot dog twisted its body on the exhibition stage, breaking into a comedic dance. Find friends, say hello, dance—this robot dog really listens. On the morning of October 20, the 2026 Physical AI and World Model Conference opened in Nanjing. Turing Award winners shared frontier judgments via video; academicians and industry guests gathered on site; a batch of innovative achievements also appeared in the exhibition area. At Agnes AI's booth, making the robot dog perform such actions previously required manual remote control. Now, users只需 say the desired result, and the rest is left to AI.
展区的工作台给出了线索，机器狗的摄像头视角、设备状态、模型发出的工具调用，以及设备正在执行的动作，都被整合在同一个界面中。原来，用户给出的是一句话，系统还要理解任务、观察环境、调用设备，再根据执行反馈推进下一步。爱思科技给出的答案，不只在模型里，还在连接模型与现实设备的执行框架中。在自研的全模态模型基础上，他们构建了一套智能体运行框架，负责连接模型、工具、任务状态与执行流程，再通过设备接入协议MHS，将这套能力延伸至物理硬件。爱思科技创始人杨通曾提到，过去的机器人更多依赖遥控，而Agnes Harness希望让设备具备更强的自主决策与协作能力，学会沟通、思考并解决问题。
The booth's workbench gives a clue—the robot dog's camera view, device status, tool calls issued by the model, and the device's executing actions are all integrated into the same interface. So the user gives one sentence; the system must understand the task, observe the environment, call devices, and advance the next step based on execution feedback. Agnes AI's answer lies not only in the model but in the execution framework connecting models and real devices. Based on its self-developed full-modal model, it built an agent operating framework responsible for connecting models, tools, task states, and execution flows, then extending this capability to physical hardware through the MHS device access protocol. Agnes AI founder Yang Tong mentioned that past robots relied more on remote control, while Agnes Harness hopes to give devices stronger autonomous decision-making and collaboration capabilities, learning to communicate, think, and solve problems.
现场一段演示里，机器狗Go2和机器人小车Car01接到了一项共同任务。一句指令，同时涉及两台设备、多个动作，以及明确的执行顺序和时长要求。从展示的工具调用记录来看，系统首先调用list_media查询音频资源，确认音乐文件存在。随后分别调用两台设备的look功能，读取摄像头画面，检查现场情况。接着，系统通过say功能安排双方用英语打招呼、回应。问候结束后，小车调用play_audio播放音乐，Go2通过action执行舞蹈动作。第一段舞蹈结束后，系统发现用户要求的30秒时长尚未满足，于是继续发起下一段动作调用。也就是说，一次工具调用结束，并不意味着整个任务结束，系统还需要对照最初的目标，检查当前进度，再决定下一步行动。
In one on-site demonstration, robot dog Go2 and robot car Car01 received a joint task. One instruction simultaneously involved two devices, multiple actions, and explicit execution order and duration requirements. From the displayed tool call records, the system first called list_media to query audio resources, confirming the music file exists. Then it separately called both devices' look function, reading camera feeds and checking on-site conditions. Next, the system used the say function to arrange both parties to greet and respond in English. After greetings, the car called play_audio to play music, and Go2 executed dance actions via action. After the first dance segment ended, the system found the user-required 30-second duration was not yet satisfied, so it continued initiating the next action call. In other words, one tool call ending does not mean the entire task ends; the system must compare against the original goal, check current progress, and then decide the next action.
Agnes Harness由大脑、小脑、反射弧、记忆与主动性、学习能力、身体组成。大脑负责想清楚要什么，面对用户需求，它要识别任务目标、梳理步骤关系，形成高层行动计划。小脑负责结构化决策，JEV模型会结合当前状态和可选行动做出决策，与大脑分工协作。反射弧负责把一连串动作串起来，复杂任务往往涉及多个步骤，有些需要按顺序执行，有些则可以并行推进。Agent Loop牢牢盯住任务，确保事情持续推进，它维护执行上下文与任务状态，记录哪些操作已经完成、工具返回了什么结果，还有哪些目标尚未达成。至于机器狗、小车、机械臂这些硬件设备，它们才是负责干活的身体，需要通过统一的接入机制，才能被脑和中枢神经发现和调用，而承担着连接不同设备工作的就是MHS。
Agnes Harness consists of brain, cerebellum, reflex arc, memory and proactivity, learning capability, and body. The brain figures out what is wanted; facing user needs, it identifies task goals, sorts out step relationships, and forms high-level action plans. The cerebellum handles structured decision-making; the JEV model combines current state and available actions to make decisions, collaborating with the brain. The reflex arc strings a series of actions together; complex tasks often involve multiple steps, some requiring sequential execution, others parallel. Agent Loop keeps a firm eye on the task, ensuring continuous progress—it maintains execution context and task state, records which operations are complete, what results tools returned, and which goals remain unmet. As for robot dogs, cars, robotic arms—these hardware devices are the bodies doing the work, requiring a unified access mechanism to be discovered and called by the brain and central nervous system; the MHS handles connecting different devices.
现场工作人员介绍，为了让整套能力有机会在一个可控环境中接受更为复杂任务的验证，爱思科技还自建了一套虚拟世界沙箱模拟系统。比如火星救援任务中，面对宇航员需要救援这样一句模糊指令，系统需要自行规划行动。先由无人机升空巡航，定位被困者并探测现场；发现氧气短缺后，再让机器狗搜寻氧气瓶，最后由运载小车将物资送往指定位置，完成补给。这里的重点，不是某一台设备能完成多少动作，而是不同设备能否围绕共同目标分工，并根据任务进展衔接下一步行动。
On-site staff introduced that to let the entire capability be validated in a controlled environment with more complex tasks, Agnes AI built its own virtual world sandbox simulation system. For example, in a Mars rescue mission, facing the vague instruction "astronaut needs rescue," the system must plan actions itself. First a drone takes off to patrol, locate the trapped person and survey the scene; after discovering oxygen shortage, the robot dog searches for oxygen tanks; finally a transport car delivers supplies to the designated location, completing resupply. The focus here is not how many actions one device can complete but whether different devices can divide labor around a common goal and connect next actions based on task progress.
任务完成了，留下的经验还能干什么。答案是能，而且是一套双进化智能系统，系统层的递归自我改进，以及模型层的测试时训练。前者改进智能体怎么干活，后者探索模型怎么从经验中学习。展台上播放的王者荣耀画面中，AI正在指挥英雄对局。一局打完之后，教练模型会读取对局记录，检查关键决策与执行结果，再尝试修改策略提示词、决策选项或相关代码，并通过后续测试检验效果。技术报告显示，在针对游戏内置Bot的测试中，1v1最佳策略的胜率从约3%提升至90%；另一组独立的3v3实验经过三轮RSI，胜率达到约60%。现场大屏显示，5v5实验经过五轮RSI，胜率已经达到100%。
After tasks are completed, what can the retained experience do. The answer is yes—a dual-evolution intelligent system, system-level recursive self-improvement, and model-level test-time training. The former improves how the agent works; the latter explores how the model learns from experience. On the booth's Honor of Kings screen, AI is commanding heroes in a match. After a game ends, the coach model reads the match record, checks key decisions and execution results, then tries modifying strategy prompts, decision options, or related code, and tests effects through subsequent matches. Technical reports show that in tests against built-in game Bots, 1v1 best strategy win rate rose from about 3% to 90%; another independent 3v3 experiment after three RSI rounds reached about 60% win rate. On-site large screens show 5v5 experiments after five RSI rounds reaching 100% win rate.
另一块大屏展示了机械臂仿真实验。画面中，机械臂反复尝试把红色方块放入目标盘。每次执行产生的动作结果与模型版本信息，都会被记录下来。演示中，Agnes Embed TTT会利用机械臂执行任务时积累的经验，调整模型中负责决策的部分参数。模型原有的主体部分保持不变，重点训练的是负责判断下一步该怎么做的决策模块。据工作人员介绍，系统会先训练模型，再评估其表现是否达到上线要求。如果达到要求，就启用更新后的模型；如果仍未达到要求，进一步优化Harness，尝试细化任务拆解，重新采集机械臂运动轨迹，再进行训练与评估。Harness的调整不仅影响任务怎么执行，还会影响后续训练数据怎么产生。
Another large screen showed robotic arm simulation experiments. In the footage, the robotic arm repeatedly tries to place red blocks into a target tray. Each execution's action results and model version information are recorded. In the demonstration, Agnes Embed TTT uses experience accumulated during robotic arm task execution to adjust some parameters in the model responsible for decision-making. The model's original main body remains unchanged; training focuses on the decision module responsible for judging "what to do next." According to staff, the system first trains the model, then evaluates whether its performance meets launch requirements. If it does, the updated model is enabled; if not, the Harness is further optimized, attempting finer task decomposition, re-collecting robotic arm motion trajectories, then training and evaluating again. Harness adjustments not only affect how tasks execute but also affect how subsequent training data is generated.
从模型到Harness，再到具体应用与开发者生态，爱思科技想打通的，是一条将模型能力转化为实际工作成果的链路。从看起来很聪明到真正能干活，这场仗不好打。但让技术走出实验室、把能力做进产品、再把应用铺进千行百业，这场硬仗，或许才是中国科技公司真正有机会打出优势的地方。
From model to Harness, to specific applications and developer ecosystem, what Agnes AI wants to connect is a chain converting model capability into actual work results. From "looks very smart" to "can truly work"—this battle is not easy. But getting technology out of the laboratory, building capability into products, and spreading applications across thousands of industries—this hard battle may be where Chinese tech companies truly have a chance to gain advantage.
