HousMthr Launches AI Travel Co-Pilot to Automate Live Group Logistics

HousMthr推出AI旅行副驾，实现团队实时后勤自动化 当静态旅行计划失效时，HousMthr利用实时AI智能体解决即时行程变故、动态费用分摊及团队安全问题。 Will Schmahl解释称，传统旅行应用侧重于行前预订，而HousMthr则致力于解决旅行中计划频繁受阻的团队执行难题。该平台采用由中央LLM统筹“社交”（Social）、“安全”（Safety）、“后勤”（Logistics）和“语境”（Context）四个专用子智能体的架构，以管理动态日程变动、自动费用分摊、本地商家优惠及紧急安全功能。通过分析长尾群聊数据，HousMthr提供前瞻性且高度契合情境的建议，以适应不断变化的团队动态。展示最后以问答环节结束，详细介绍了团队定义、自驾后勤、每年12.99美元的订阅模式以及B2B商家合作关系。 HousMthr创始人兼CTO Will Schmahl介绍了该平台如何通过中央LLM协调社交（Social）、安全（Safety）、物流（Logistics）与语境（Context）四大子Agent，专门应对旅行开始后的突发混乱与行程变动。 (0:04) 系统追踪279个涵盖天气、安全与预订变更的实时触发场景，并直接嵌入共享日历和记账工具，可自动完成因成员生病缺席而将其移出账单分摊等行政协调事务。 (3:08) 在架构上，平台为每趟行程单独启动LLM实例并接入目的地知识库进行即兴协同，同时配备语音AI功能，可在紧急情况下向群组广播定位或直接连通911报警求助。 (5:08) 通过分析行前长达数月的群聊长尾数据，HousMthr可精准挖掘偏好并提供即时建议；产品面向两人以上团体及拼房单人旅客，计划以每年12.99美元订阅费及B2B本地商家尾单集市实现变现。 (9:02)

Will Schmahl, Founder and CTO of HousMthr Inc., argues that existing travel AI focuses excessively on pre-trip booking while ignoring the chaotic, real-time friction that occurs once group trips begin. HousMthr deploys a multi-agent AI architecture to manage live travel logistics, safety interventions, and social dynamics dynamically. Targeting In-Stay Logistics Over Booking (0:04) Schmahl identifies a structural gap in travel tech: after booking, travel groups are left to navigate dynamic disruptions independently, placing heavy administrative burdens on an informal group organizer. During live trips, small issues—such as missed flights, delayed cabs, or lost groceries—often trigger cascading disruptions. Coordinating Specialized Sub-Agents (5:08) To handle changing group moods and environmental shifts, HousMthr utilizes a unified intelligence framework running individual trip-level LLMs underneath core destination models, which Schmahl refers to as a 'bible of triggers.' This system orchestrates four dedicated sub-agents: Social, Safety, Logistics, and Context. Interacting like an improvisational jazz band, the agents automate collaborative workflows such as shared calendars and bill splitting; for example, the system automatically removes an individual from a restaurant bill split if they skip dinner due to illness. Resolving Live Triggers and Safety Risks (5:08) HousMthr tracks 279 distinct use cases for real-time triggers spanning weather, bookings, and safety. Powered by member-shared mobile location data, the platform provides notifications ranging from localized photo opportunities to severe weather alerts for thunderstorms or tornadoes. In urgent scenarios, audio-based agents can broadcast location updates to group members or patch directly into local 911 emergency services. Leveraging Long-Tail Pre-Trip Conversations (7:38) Traditional platforms lose visibility between booking and arrival, but HousMthr's pilot captured up to nine months of ongoing chat data while users coordinated seasonal rentals and expenses. By analyzing these long-tail exchanges for personal preferences, such as music tastes, the system delivers tailored, real-time recommendations when groups reach their destination, contrasting static itineraries with adaptive plans. Group Scope, Monetization, and Local Marketplaces (9:02) HousMthr defines a group as two or more travelers, with solo travelers able to match into groups with spare rooms, and supports edge cases like multi-car road trips by monitoring arrival times to avert logistics failures. The company plans to monetize through a $12.99 annual subscription for unlimited trips—transitioning from a free pilot—and a B2B marketplace where local businesses offer discounted last-minute inventory to fill unsold capacity. HousMthr aims to stabilize group travel by continuously adapting to live operational friction, though its model depends on user adoption of real-time location sharing and direct integration with local merchants.

HousMthr公司创始人兼首席技术官Will Schmahl指出，现有多数AI旅游工具局限于行前预订，忽视了行程开启后的实际应对混乱；该公司通过构建由多智能体协同驱动的统一群体智能平台，旨在实时化解多人出行中的突发变故、后勤分工与安全风险。 行中即时履约与多人旅行痛点 (0:04) Schmahl在演讲中表示，多数旅游AI产品将重心完全置于行前的行程规划与票务预订，但旅客完成预订后往往只能独自面对复杂的实际度假过程。多人团队出行呈现高度的动态特征，组员的个人偏好、情绪状态及外部环境在旅途中持续演变。在亲友组成的旅行团队中，通常存在一位承担全部统筹压力的“核心组织者”（即注重秩序的Type-A型规划者），该角色亟须一种能够向全体成员分担后勤职责的协同机制。 基于为期一年的试点观察，HousMthr将产品切入点定位于“停留期”或“行程中”（in-stay / in-travel）的具体执行环节。多人旅行常因错过出租车、航班延误、未赶上轮渡或食物短缺等不可预见的变故产生“滚雪球效应”，甚至导致原定的静态计划全面瓦解。平台试图通过自适应AI智能体实时响应环境变化，防止局部突发状况破坏整体行程。 统一群体智能与多智能体分层架构 (5:08) 为应对多变的群体动态，HousMthr设计了以“统一群体智能大语言模型”（Unified Group Intelligence LLM）为核心的中枢架构。该中枢统筹调度四个专职子智能体：负责群体互动与氛围的社交智能体（Social Agent）、负责预警与求助的安全智能体（Safety Agent）、负责物资调配与行程调整的后勤智能体（Logistics Agent），以及负责环境与位置感知的情境智能体（Context Agent）。Schmahl将这些智能体间的交互协作比作“爵士乐队”，强调其能根据现场数据流展开即兴调整与流体化响应。 在底层技术设施上，平台会针对单次行程单独实例化专属LLM，其底层依托存储有本地化知识库（被称为“触发规则圣经”）的目的地核心LLM。该架构旨在深耕行中体验而非依赖传统的票务佣金分润。例如，系统若检测到群体傍晚聚集于海滩度假屋，会在一小时后主动推送交互游戏促进互动；在成员失联或遭遇危险时，系统可通过音频AI智能体广播定位，甚至直接接入当地911紧急救援热线。 情境触发用例与协作工具自动化 (5:08) HousMthr将AI智能体直接内嵌至共享日历、多方记账和活动投票等协作模块中，利用情境逻辑消除群体内部的行政管理摩擦。以用餐记账为例，若系统检测到某位成员因身体不适未参与餐厅晚宴，会自动在费用分摊时将其剔除，无需组员人工核对。其AI风险处置机制能够持续处理现场情境触发事件，自主评估突发状况、制定解决策略，并主动向团队执行或确认方案。 据Schmahl介绍，该系统目前涵盖了279个涉及安全防范、气象突变与预订变更的独立实时触发用例。借助团队成员授权共享的移动端实时地理位置数据，系统兼顾游玩灵感与风险预警：既能在两街区外出现优质拍照点时发送提示，也能在面临雷暴或龙卷风预警等恶劣天气时推送防御通知，构筑全方位的防护与支持网络。 长尾数据沉淀与双向商业变现 (7:38) Schmahl指出，传统旅游平台无法留存用户行前的大量深层交流数据。HousMthr在度假短租试点期间，沉淀了用户长达九个月的行前群聊与支付沟通记录。通过分析这些长尾数据，系统能挖掘用户的细分兴趣（如特定音乐偏好或活动类型），并在旅途中一旦出现匹配资源时实施超本地化精准推荐。在适用场景方面，系统将团队定义为两人及以上，单人旅客亦可借此匹配闲置床位；在多车自驾场景中，系统能基于位置预测成员接驳情况（如判断是否会错过末班轮渡），提前提醒就近队友采购生活物资。 商业模式上，HousMthr在试点阶段保持免费，并计划在演讲后三周内上线完整AI功能套件。C端用户采用每年12.99美元的无限制出行订阅制；B2B端则搭建了目的地商户集市，允许当地餐厅等商户临期折价投放闲置名额（如为10人团队提供30%折扣），借此实现商家盘活闲置资源与旅客享受团购优惠的双向互利。 Schmahl展示了HousMthr将AI介入点从行前预订拓展至行中履约的自适应协作构想，但该方案高度依赖全员实时位置共享与长期社交聊天挖掘，其在大规模推广中的隐私边界保障、复杂网络环境下的智能体鲁棒性，以及12.99美元低价年费对算力成本的支撑能力，仍有待后续实际检验。

## Background 本内容由频道 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布，记录了 HousMthr 创始人兼 CTO Will Schmahl 的主题演讲。根据页面介绍，演讲主要探讨了 HousMthr 如何利用 Agentic AI 解决群组旅行在“行程中”面临的突发混乱与协同痛点，分享了多 Agent 架构设计、实时环境适应以及消费级 AI 从行前规划转向现场执行的核心经验。 ## TL;DR Will Schmahl 介绍了专为两人及以上群组出行打造的 AI 实时协同平台 HousMthr。针对传统旅游工具只管行前预订、缺乏行中支持的痛点，团队构建了由社交、安全、后勤和上下文四大 Agent 组成的“统一群体智能”架构，并内置 279 种实时触发器来前瞻性化解误车、恶劣天气或分账繁琐等突发问题。平台提出如爵士乐队般即兴协同的交互模式，并通过 12.99 美元/年的亲民 C 端订阅与连接目的地商户闲置库存的动态折扣撮合，实现从行程守护到商业变现的闭环。 ## Atmosphere & style 整场演讲呈现出典型硅谷创业者务实、生动且充满激情的分享风格。Will Schmahl 开场以轻松幽默的自嘲拉近与观众的距离，随后结合自己在海岛错失轮渡等真实生活窘境展开论述，叙事既接地气又极具代入感。现场氛围专注且互动热烈，演讲者将复杂的双层 LLM 与多 Agent 架构比喻为行云流水的爵士乐即兴演奏，生动传递出“让自适应智能体化解静态计划崩溃”的技术信心与人情关怀。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=4s) 传统在线旅游平台（OTA）大多在完成订单后便结束服务，群组出行的真正痛点集中在充满不确定性的“行程中”，且往往导致单一组织者承受过重负担。 - [01:15](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=75s) HousMthr 构建了由社交、安全、后勤和上下文四大专属 Agent 组成的协同架构，在中央 LLM 统领下实现针对单次旅程的“统一群体智能”。 - [03:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=184s) 系统内置 279 种多人出行“实时触发器”，涵盖突发天气、延误预警以及动态分摊账单（如自动剔除未就餐人员）等场景的前瞻性干预。 - [05:08](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=308s) 多 Agent 协同如同“爵士乐队即兴演奏”，能敏锐捕捉群组成员的行程空隙与情绪状态，主动推送适配的互动游戏或本地体验。 - [05:08](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=308s) 技术上采用双层 LLM 体系，为每个出行群实例化专属临时 LLM，并叠加基于目的地长期沉淀的核心 LLM，同时支持精准语音求助与紧急报警联动。 - [08:47](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=527s) 产品的核心理念是“静态计划往往崩溃，自适应智能体随之应变”，强调 AI 应转型为旅途中的实时副驾驶而非单纯的行前攻略工具。 - [08:47](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=527s) 平台将群体定义为两人及以上，并针对独行旅客提供偏好匹配机制，协助其拼入拥有空闲床位且节奏合拍的旅行团队中。 - [09:54](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=594s) 针对多车结伴自驾（caravanning）场景，系统通过成员位置与时空预判算法，可在某车延误或加油时主动向全队广播并推送替代应对方案。 - [11:07](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=667s) 商业化路径采用 12.99 美元/年的 C 端平价订阅制，结合 B 端目的地商户闲置库存（如餐厅大桌与活动尾货）的高折扣即时撮合实现双赢。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=4s) HousMthr 的诞生与群组出行的痛点 演讲开场时，Will Schmahl 带着轻松调侃的语气向观众致意，笑称台下喜欢旅行的人似乎比预想的要少，“这倒挺奇怪的”。他直言，这场演讲与过去三天大家在大会上看到的任何内容都截然不同——他是以创业者和 CTO 的身份站在这里，展示团队如何将前沿技术落地为初创产品 HousMthr，并探索如何借助 AI 变革人们的现实生活体验。 HousMthr 创立的核心目标，是协助群体旅客管理“行程中”（in-stay / in-travel）的复杂需求。现有的在线旅游平台几乎都在用户完成预订或拿到确认单那一刻戛然而止，后续与朋友或家人的旅程如何落地，全部留给旅客自行摸索。在大多数亲友出行中，总会有一个充当“领头人”的 A 型人格规划师（Type A personality）。这个人往往承受着巨大的心理包袱，竭力确保每个人都玩得开心；但一段健康的群组旅程，理应把责任分散开来，让每个人都承担相应的角色。在过去一年的试运行（pilot）中，团队敏锐地意识到，借助 AI 能够打造出强大的实时工具，大幅推进旅行中的奖励激励与安全防护机制。 ### [01:15](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=75s) 现场混乱与四大 Agent 协同架构 群组旅行往往伴随着现场的动态混乱。只要出现一个纰漏——某人没赶上出租车、有人误了航班或轮渡、有人突然没饭吃——就会引发连锁的滚雪球效应（snowball effect）。群组出行的痛点在于群体动态时刻在变：每个人的性格、兴趣、即时偏好各不相同，既有个人意愿，又有全团共同想做的事，甚至还取决于大家当天的情绪状态。如果没有快速响应的 AI 系统，去实时协调这些变化极其困难。 为了应对这些输入并产生合理的行动输出，团队在底层系统架构中构建了四大各司其职的智能体（计划于两周后正式发布）： - **社交智能体（Social Agent）**：负责照顾群内成员之间的社交氛围、互动与关系维系。 - **安全智能体（Safety Agent）**：实时监测各类安全警报与风险事件。 - **后勤智能体（Logistics Agent）**：负责行程变动、交通换乘与日常落地协调。 - **上下文智能体（Context Agent）**：捕获当前所处地理环境、天气及行程所处的时间节点。 这四大 Agent 既相互独立运作，又在一个中央大语言模型（LLM）系统的统领下紧密协作，构成了针对单次旅程的“统一群体智能”（Unified Group Intelligence）。当用户将目的地录入系统后，后台便会调取该地的历史数据，主动提示潜在风险，在确保安全的同时为行程增添惊喜。 ```mermaid flowchart TD accTitle: HousMthr 多 Agent 架构图 accDescription: 描述用户输入通过中央大语言模型系统分配给四大 Agent 协同处理的架构 subgraph Inputs["输入层与工具集"] direction TB I1["群组偏好与行程数据"] I2["日历 / 账单 / 地理位置"] end subgraph Brain["核心决策层"] LLM["统一群体智能 (中央 LLM)"] end subgraph Agents["四大专属智能体"] direction TB A1["社交智能体 (Social)"] A2["安全智能体 (Safety)"] A3["后勤智能体 (Logistics)"] A4["上下文智能体 (Context)"] end subgraph Execution["输出与反馈"] O1["主动安全警报"] O2["动态后勤调整"] O3["体验与奖励推荐"] end Inputs --> Brain Brain --> Agents Agents --> Execution ``` ### [03:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=184s) 279 种实时触发器与风险管理机制 HousMthr 为群体旅客提供了一整套实用工具箱，涵盖共享日历、分摊记账、活动选择、房间分配以及预算上限控制等。四大智能体直接接入这些工具中进行自动化优化。Will Schmahl 举了一个生动的场景：如果全团晚上外出聚餐，但其中有一位成员因为身体不适留在住处未参加，当大家录入账单平摊费用时，系统会自动将未到场的那名成员剔除，避免了事后核对算账的尴尬与繁琐。 在系统底层，团队梳理并建立了多达 279 种针对两人及以上群组出行的“实时触发器”（Live Triggers），涵盖安全突发、恶劣天气、预订变动等方方面面： - **触发与评估阶段**：系统实时捕获异常（例如龙卷风预警或行程延误），迅速评估其对群组成员的影响。 - **执行与确认阶段**：系统采取应对行动，主动向用户发出预警，并与团队确认解决方案。 - **双向赋能机制**：这套机制不仅用于化解风险，同样用于挖掘奖励——小到主动提醒“距离你两个街区外就是绝佳拍照打卡点”，大到针对突发特大雷暴或龙卷风预警发出严肃避险通知。 只要用户自愿共享移动端数据，系统就能在旅客甚至尚未察觉周围状况之前，完成前瞻性的预判与干预，大幅减少因突发事件导致的行程延误和挫败感。 ### [05:08](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=308s) 爵士乐队式协同：即时体验驱动与商业闭环 与传统在线旅游巨头死磕“行前机票酒店销售”（pre-trip sales）截然不同，HousMthr 的重心全放在“正在旅行中”的群体及其体验优化上。虽然产品未来也会向用户推荐消费项目，但形式将完全基于当日折扣（day sales），并将流量动态引导至目的地当地的实体商家与社区（Destination Communities）。 Will Schmahl 将团队的多 Agent 协同模式比喻为一支“爵士乐队”：“他们就像爵士乐手一样在即兴演奏（improvising）”。Agent 会根据群组动态的演进，寻找恰当的缝隙与用户互动。例如，系统监测到全员正在海滩别墅度假，晚上整组人都在住处共进晚餐；在他们用餐完毕一小时后，社交智能体便会主动推送一款适合餐后全团参与的互动小游戏。 在技术架构与安全方面： - **安全与语音求助**：若团队成员在陌生环境中走散，用户可直接通过语音智能体触发求助，告知其当前精确方位，甚至能在紧急情况下直接转接当地 911 报警服务。 - **双层 LLM 体系**：团队正在攻关为每一个出行群独立实例化一个专属的临时 LLM 实例；在这之上，则部署基于目的地长期沉淀的“核心目的地 LLM”，作为整个规则与触发器库的底层基石。 - **私域数据商业化**：在季节性度假租赁（seasonal rentals）试点中，许多群组成员为了订下雪季或旺季的房子，提前 9 个月就在应用内筹备群聊。系统在脱敏分析中掌握了他们喜爱的音乐家或活动喜好。当他们到达目的地且恰逢该艺术家巡演时，系统便能顺理成章地向其推荐门票或定制活动，自然跑通商业化变现。 ### [08:47](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=527s) 核心理念与现场问答：静态计划崩溃，自适应智能体应变 Will Schmahl 总结道，传统做法的最大弊端在于“静态计划往往会崩溃，而自适应智能体能够随之调整”（"Static plans collapse and adaptive agents adapt"）。AI 的角色绝不仅是行前做攻略，而是成为与团队同行的实时副驾驶，通过对环境、情绪与现场突发情况的主动响应，让在外奔波的群体旅客感受到像家一样的关怀——这正是“HousMthr”（管家/宿母）名字的由来。 在随后的现场问答环节中，观众提出了一个基础但关键的定义问题： - **问：你们如何定义一个“群体”（Group）？** Will Schmahl 回答：两人及以上（two plus more）即构成群体。 - **补充说明（独行旅客机制）**：单人旅客（solo traveler）同样可以使用该平台，但其核心逻辑在于匹配机制——系统会帮助独行旅客寻找与其旅行节奏、偏好契合的现有群组；若该群组恰好有空闲的床位或房间，便能将其吸纳进来，合并成为新旅程的一分子。 ### [09:54](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=594s) 多车结伴自驾与实时动态协同 在问答环节，有观众提问 HousMthr 的系统是否考虑了公路旅行以及多辆车结伴车队出行（caravanning）的场景。Will Schmahl 肯定地表示，团队正在构建的功能非常契合这类需求，而这个产品灵感正源于他的一次亲身经历。当时他和朋友们在海岛上租了一栋海滨别墅，其中一人负责采购全团的食物，结果那个人错过了最后一班轮渡，而岛上的本地商店当晚也即将打烊。虽然他们最终靠着岛上的熟人解决了温饱，但这促使他们将该场景作为典型用例落地到产品中。 在 HousMthr 的实时协同设计中，AI Agent 能基于地理位置主动计算并预判出“某人注定赶不上末班轮渡”。与此同时，AI 会自动向岛上其他成员发出预警通知：“当地商店还有 20 分钟关门，赶紧去买东西，因为 Roger 赶不上这班船了。”在车队自驾游场景下，这种机制同样适用——系统不仅能支持拼车协调，还能追踪各车辆的实时动态；一旦前车或后车停下来加油，系统便能及时通知同行者“他们在加油，稍等一下”，从而解决旅途中信息沟通滞后、缺乏主动前置监控的痛点。 ```mermaid flowchart TD accTitle: HousMthr 实时多车协同与突发事件预警流程 subgraph S1["位置与动态监测"] A["多车/成员实时位置与行进轨迹"] --> B["AI 实时时空预判 Agent"] end subgraph S2["突发与异常检测"] B --> C{"检测到延误风险<br/>(如错过末班轮渡/临时加油)"} end subgraph S3["上下文协同与主动决策"] C -->|触发补救行动| D["检索周边环境信息<br/>(如岛上商店 20 分钟后闭店)"] D --> E["向在场同伴推送紧急操作建议<br/>'Roger 赶不上船，立即前往商店采购'"] C -->|自驾协同| F["向车队同步状态<br/>'后车正在加油，请靠边等待'"] end ``` - **自驾与车队痛点**：传统群聊和电话在驾车过程中无法实现高频、主动的多方步调对齐，信息往往“传递得够快，但不够协调，也缺乏前置主动监测”。 - **主动干预逻辑**：基于成员空间轨迹与外部营业时间/交通班次进行多源交叉分析，在危机发生前将替代方案派发给具备执行条件的同伴。 --- ### [11:07](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=667s) 商业化路径：C 端订阅与 B 端尾货动态撮合 针对现场关于盈利模式（收入是来自 C 端用户还是 B 端商户/活动推荐）的提问，Will Schmahl 阐明了 HousMthr 当前的商业化路线。目前平台处于试点阶段尾声，用户可以免费注册使用基础的行程搭建、房屋管理与好友资源整合等工具集；而全套 AI Agent 核心功能将在三周后正式上线。 在收费模式上，面向消费者的订阅定价为极其亲民的 **12.99 美元/年**，用户可享受全年在平台上组织无限次旅行的权限。除此之外，HousMthr 正在探索重塑目的地商家的交易生态：通过实时连接旅行团体的动态画像与目的地商户的闲置库存，实现精准需求撮合。例如，当一家餐厅临时有一张 10 人的空大桌时，系统可以将该名额精准推送给附近刚好有 10 人的旅行团体，前提是商家必须提供深度折扣（如 30% off）。这种模式既为线下商户填补了未售出的闲置库存，又让旅行者获得了切实的实惠，构建起双赢的商业闭环。 - **C 端定价**：年费 12.99 美元，支持无限次团体出行规划与实时 AI 协同。 - **产品上线节奏**：基础工具箱已开放免费体验，AI Agent 核心功能在演示后 3 周内全量上线。 - **B 端动态分销**：利用群组规模与实时位置，以“深折扣（如七折）换即时闲置库存核销”的方式切入目的地餐饮与活动市场。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=4s) 这倒挺奇怪的 — Will Schmahl > [03:04](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=184s) 距离你两个街区外就是绝佳拍照打卡点 — Will Schmahl > [05:08](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=308s) 他们就像爵士乐手一样在即兴演奏（improvising） — Will Schmahl > [08:47](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=527s) 静态计划往往会崩溃，而自适应智能体能够随之调整（Static plans collapse and adaptive agents adapt） — Will Schmahl > [08:47](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=527s) 两人及以上（two plus more）即构成群体 — Will Schmahl > [09:54](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=594s) 当地商店还有 20 分钟关门，赶紧去买东西，因为 Roger 赶不上这班船了。 — Will Schmahl > [09:54](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=594s) 他们在加油，稍等一下 — Will Schmahl > [09:54](https://www.youtube.com/watch?v=lC_1LVlTj6Y&t=594s) 传递得够快，但不够协调，也缺乏前置主动监测 — Will Schmahl ## Mentioned Will Schmahl, HousMthr Inc., HousMthr, Ai4, 社交智能体 (Social Agent), 安全智能体 (Safety Agent), 后勤智能体 (Logistics Agent), 上下文智能体 (Context Agent), 中央大语言模型 (Central LLM), 911 报警服务 ## Video description Will Schmahl, Founder and CTO, HousMthr Inc. Abstract What if AI agents weren’t just assistants before a trip—but became co-pilots during it? I’ll explain how HousMthr is using agentic AI to solve the chaos of group travel once the trip begins—when plans shift, people scatter, and the itinerary breaks. This talk will illustrate how we designed AI agents that handle everything from dynamic logistics and safety alerts to real-time group coordination and contextual local suggestions. I’ll share key lessons around building for live environments, designing agents for multi-user group dynamics, and why execution—not planning—is the next big opportunity for consumer AI. If you're working on AI tools that need to operate in unpredictable, real-world settings, this session will give you a blueprint. [Source](https://www.youtube.com/watch?v=lC_1LVlTj6Y)

[0:04] Will Schmahl: 这倒挺奇怪的 [3:04] Will Schmahl: 距离你两个街区外就是绝佳拍照打卡点 [5:08] Will Schmahl: 他们就像爵士乐手一样在即兴演奏（improvising） [8:47] Will Schmahl: 静态计划往往会崩溃，而自适应智能体能够随之调整（Static plans collapse and adaptive agents adapt） [8:47] Will Schmahl: 两人及以上（two plus more）即构成群体 [9:54] Will Schmahl: 当地商店还有 20 分钟关门，赶紧去买东西，因为 Roger 赶不上这班船了。 [9:54] Will Schmahl: 他们在加油，稍等一下 [9:54] Will Schmahl: 传递得够快，但不够协调，也缺乏前置主动监测

[0:04] Will Schmahl: snowball effect [0:04] Will Schmahl: ring leader [5:08] Will Schmahl: jazz band [5:08] Will Schmahl: the bible of triggers [7:38] Will Schmahl: static plans collapse [9:02] Will Schmahl: two plus more

[0:04] Thanks, guys. Surprised now there's not [0:05] more of you that like travel. It's kind [0:07] of weird. Um, this is a little bit [0:09] different than anything else you've seen [0:10] probably the last three days. Um, I'm [0:13] coming here as a founder showing what [0:15] how we're using AI. Um, and how we're [0:18] looking to actually bring the world into [0:20] a different place with the advances that [0:22] we have uh come to with just technology [0:25] in general uh focused on you know our [0:27] startup called Housemother. So to give [0:29] you a little bit of um you know uh [0:31] insight, so Housemother was created to [0:33] help groups manage their instay or in [0:36] travel needs. So most of the time you [0:38] get your booking done and you're giving [0:39] your reservation and then you're off to [0:41] figure out how you handle that trip or [0:43] that you know u vacation with your [0:45] friends or family. But there's always [0:47] that planner that lives amongst friends [0:48] and family groups that is the ring [0:51] leader, the person a type personality [0:53] that focuses on making sure that [0:56] everyone's happy. But you still need to [0:58] spread the wealth and make sure that [0:59] everyone has some responsibility. So [1:01] that's why Housemother was created. And [1:02] as we did a pilot over the last year, we [1:05] realized that you know with AI we could [1:07] really empower tools that could you know [1:09] bring advances into reward and safety [1:11] mechanisms which I'm going to go over [1:12] today. [1:15] So uh you know there's planning and [1:16] chaos you know amongst groups. It's real [1:19] time. It happens live. There's always, [1:22] you know, someone misses a cab ride, you [1:24] miss a a flight, you miss a ferry, you [1:26] know, people don't have food. There's [1:28] there's different sequences that happen. [1:30] It's a snowball effect, you know, as you [1:32] go along traveling with friends or [1:33] family. So, you know, in context, what [1:36] we look at, you know, the system of at [1:38] hand is that we have different calibers [1:40] of data that we're reading that go from [1:43] uh basically the inputs, the outputs. In [1:46] the middle, we have uh created four [1:47] different agents in the middle which [1:48] we're releasing in about two weeks. And [1:51] as we look at the data, it's quite [1:52] interesting the way we looked at it. So [1:55] we have initially um you know what makes [1:57] it hard is that share groups the they [2:00] have dynamics that change personalities [2:02] interest preferences right you have [2:04] individual preferences and then you have [2:05] group preferences of what people want to [2:07] do you know how people are feeling that [2:09] day um all those factors come into play [2:12] and it's hard without you know fast [2:14] systems like AI. So also you have [2:16] decisions uh that change and also just [2:18] you know environmental uh and contextual [2:20] uh changes and effects that happen to [2:23] the group. [2:25] So the four categories that we have [2:26] actually worked on is a social agent uh [2:29] safety agent uh a logistics agent and [2:32] then a context agent. Now these all work [2:34] independently but then cohesively below [2:37] a central you know LLM LLM system that [2:41] basically uh works together as a major [2:44] this one group is going on one trip. We [2:46] know what the destination is. They put [2:48] that trip into the system and from there [2:50] we go back to data looking at what has [2:53] happened at that location what we maybe [2:55] have to proactively let the users know [2:57] about and how we can keep their trip [2:59] better and give them rewards and also [3:01] keep them safe. [3:04] So this is under the the uh unified [3:06] group intelligence I was just speaking [3:07] about. So you have different room [3:09] assignments, you know, you have budget [3:11] caps, you have things that are important [3:12] to users as they're traveling as a [3:14] group. So what we looked at is we give [3:16] them a a suite of tools which is [3:18] everything from shared calendar, shared [3:20] expenses, uh events, selection of of [3:24] different things they would like to do. [3:25] And what we do is we plug in the a the [3:28] agents into those different uh groups of [3:30] tools and optimizes them so that the [3:32] groups say if you're out to dinner and [3:34] one person didn't go because they [3:35] weren't feeling well and you put that [3:36] into the expense for your group uh to [3:39] split for the trip that one person that [3:41] wasn't at that location will not be [3:43] included in the expenses. So it makes it [3:44] easier on little things that count to [3:46] make your trip better and more [3:48] manageable in a group setting. So live [3:51] triggers. So we have safety triggers, [3:52] weather triggers, booking triggers, and [3:54] then we have there's a the list goes [3:56] down to use cases of 279 we have so far [3:59] of different types of triggers that [4:01] could affect a group two plus more [4:03] people on a trip, which you know [4:04] obviously we all love to travel. Uh I'm [4:07] assuming you would all agree, but we [4:09] know that little things that come up [4:11] could be uh massive changes to our to be [4:14] delayed, to not be happy, and also not [4:16] to be satisfied as we go along our way. [4:20] So the riskmanagement agent flow so the [4:22] triggers we take a trigger we evaluate [4:25] it we then uh look at how we could act [4:28] on it and then we proactively or uh we [4:31] confirm resolution with the group. So [4:33] we're trying to make sure that not only [4:35] in a risk and reward mitigation um [4:37] mechanism. We're telling people, hey, [4:39] you're two blocks from the best place to [4:41] take a picture or there is a emergency [4:43] notification that there is a [4:45] thunderstorm, you know, that's very [4:47] massive, the tornado warning, you should [4:49] not be in that location. A lot of this [4:50] is standard, but some of it is unique in [4:52] the way of trying to proactively before [4:56] the users actually know where they're [4:57] going to be. If we're using the mobile [4:59] data that's that we're collecting, you [5:01] know, from the group, as long as they're [5:02] sharing that information with us, we can [5:04] then utilize it to make their trip [5:06] enhance. [5:08] So, creating moments. So, you know, [5:11] really what we're about is rewarding [5:13] through the agent flow moments that are [5:15] better for groups versus just using [5:18] agents to optimize sales, right? So the [5:22] travel ticket industry has been so [5:23] focused on pre-trip sales and our focus [5:27] is 100% on the group that's traveling [5:29] and optimizing their trip. Yes, we will [5:32] be selling things to users but in a way [5:35] that is dynamic through day sales that [5:37] are discounted, you know, connecting to [5:39] the local communities that are in those [5:41] destinations and those are destination [5:43] destination communities is what we call [5:45] them. [5:47] So agent communication layers. So, our [5:49] agents work as a jazz band, per se. So, [5:53] you know, they're they're improving. [5:55] They're making sure that as group [5:57] dynamics come up, they read the data. [5:59] They take chances to actually interact [6:02] with users. That means that if we know [6:04] that you're at a beach house for the [6:06] weekend and we can see that the whole [6:08] group is at one location and we know [6:10] that it's at your house and it's around [6:11] a dinner time per se, then we'll after [6:15] an hour after that being that location, [6:16] we'll then, you know, send that group a [6:19] game, right, to play after dinner. So, [6:21] we're interacting with groups to enhance [6:23] and also communicate with them to be [6:25] proactive and actually uh, you know, in [6:28] the shared knowledge that they're [6:29] sharing with us and giving it back to [6:30] them. [6:33] So, use case uh simulations. So, safety [6:35] alerts, logistic adjustments, and [6:38] experience enhancements. So, these are [6:40] standard, but what we're getting with [6:42] safety is kind of unique in being able [6:45] to actually make sure that if someone's [6:47] missing from your group, you could [6:49] automatically SLS the app via uh AI [6:52] agents that are audiobased to let them [6:54] know that you're at this location and [6:56] they should come to you. or if you want [6:58] to patch in local 911 destinations, they [7:01] could then come to you also. [7:04] So how we built this? So we have [7:06] real-time uh infrastructure, we have uh [7:09] data ingestions and then we have uh [7:11] intelligence layers. So right now um [7:13] we're still working on uh you know [7:15] individual LLMs per group, right? That [7:18] are spun up as we get the destination. [7:20] We'll then have above that actually core [7:23] uh LLMs that are based on destination so [7:25] that we could have our bible of things [7:27] or triggers that we would build off of [7:29] and then from there we're actually able [7:31] to create a travel flow per each new [7:33] trip. [7:35] Excuse me. [7:38] So the big takeaway is that uh AI isn't [7:40] planning the trip but we're it's [7:41] actually with you. So what we're trying [7:43] to do is by uh being adaptive to [7:46] actually the trip the environment and [7:48] the users at the time live we're able to [7:51] create group intelligence that actually [7:52] advances the trip keeps them safer [7:56] rewards them and it's able to actually [7:58] uh connect with them better. So by [8:00] understanding the data, the personal [8:02] preferences, the shared chat, what [8:04] they're talking about, you know, a lot [8:06] of data that we're collecting has never [8:07] been seen before because if you know [8:09] randomly, our first focus of our launch [8:11] has been a pilot in seasonal rentals, [8:13] right? Those rentals that they have, [8:15] they pay the opposite season. So for [8:17] nine months out of the year, they've [8:19] been talking in the chat, collecting, [8:21] you know, paying for their house for the [8:23] coming season. That data we've been [8:25] collecting, they've been talking about [8:26] other things in the topics of concerts [8:28] and what have you. But now we have that [8:30] data so that if they're traveling and we [8:32] have those that information to then oh [8:34] they happen to be this place and now [8:35] this artist is playing we can now [8:36] connect that into and serve that to them [8:38] as a you know option or reward. So and [8:41] that's you know then comes into our [8:43] monetization uh how we actually make [8:45] money. [8:47] So uh you know basically static plans [8:49] collapse and uh adaptive agents adapt in [8:53] our system called housemother which is [8:55] basically the concept of taking care of [8:57] a group and um making it feel like home. [8:59] That's what housemother is about. So [9:02] I'll pause there and does anyone have [9:03] any questions? [9:07] Is anyone in travel per se here? [9:10] No. Okay. [9:16] Okay. Thank you. Oh, go ahead. [9:19] >> Hold on. I have a mic for you so [9:20] everyone can hear it. [9:23] >> A simple question. What defines a group? [9:26] >> Uh, two plus more. [9:30] >> A solo traveler can use it, but uh there [9:33] the concept of a solo traveler is that [9:35] they could find a matching group to [9:38] become part of their group if they have [9:40] a say room or a bed available in the [9:43] type of travel that they're looking for. [9:45] So [9:49] >> yes, keep your hand up. [9:54] >> Uh does this include like considerations [9:56] for road trips and stuff as well like if [9:58] you have multiple vehicles traveling [9:59] together caravanning? [10:00] >> So it does actually what we're building [10:02] is kind of interesting. So what we have [10:04] is and this came out of a concept of uh [10:07] I had a we had rented my friends and I [10:10] actually a beach house that was on our [10:11] island. One person was tasked with [10:13] bringing all the food. they mess they [10:15] missed the last ferry, right? So, and [10:18] that night the the store was already [10:20] closed. So we, you know, did know people [10:22] on the island who got food. But, uh, in [10:25] the concept of you that use case, we've [10:27] already implemented where AI would have [10:30] picked up by location that that person [10:32] was not going to miss that they were [10:33] going to miss the ferry. They we would [10:35] have notified where we were in our [10:37] locations on that island that hey, the p [10:39] the the local, you know, store is going [10:41] to close in 20 minutes. Go there because [10:42] Roger's not going to make that ferry. [10:44] So, we've already taken care of [10:46] scenarios like that. So they could [10:47] actually not only carpool but you could [10:49] actually you know follow each other [10:50] figure out oh they stopped for gas let's [10:52] just wait up right so because that [10:54] communication's lost you're not you know [10:55] it's just not fast it is fast enough but [10:57] it's not coordinated enough or pro [10:59] proactively monitored that's how we're [11:01] doing it. Yeah, [11:07] >> sorry. Ju just a question about the um [11:10] house mother. Um how do you get the [11:13] payments? Uh is is it from the uh users [11:16] or from the uh events or locations that [11:19] you recommend? [11:20] >> You mean monetization or the how do we [11:21] make money? [11:22] >> How do you make money from [11:23] >> the money? So right now we're completely [11:24] free. Um and because we're still in [11:26] we're ending our pilot. Um you can sign [11:28] up today. uh the AI piece of it that's [11:31] all coming out in three weeks. Um so [11:33] right now we have all the tool sets that [11:34] are in there that basically you could [11:36] you know build your house, you could [11:38] build your trip, add your friends or [11:39] family and have all the resources. All [11:41] the AI stuff is coming 3 weeks but we're [11:43] charging $12.99 a year right for [11:45] unlimited amount of trips. monetization [11:48] then will come thereafter as far as like [11:50] obviously the the normal app you know [11:52] stuff. But we're we're also looking to [11:55] change the model of offering deals to [11:59] destination businesses uh to offer deals [12:02] to uh you know on trip so that if they [12:05] have a 10 table you know empty at a [12:07] restaurant they could offer that to a 10 [12:10] person you know that matches in that you [12:12] know scenario and then but they have to [12:14] give them 30% off. So, we're trying to, [12:16] you know, fill inventory that's [12:18] unfilled, but also give discounts to [12:20] kind of change the make it better for [12:22] travelers also. So, that's kind of our [12:24] play.