Meta, Databricks Deploy AI Swarms to Run Code and Executive Decisions

Meta与Databricks部署AI智能体集群运行代码并辅助高管决策 自主多智能体网络正将企业瓶颈从底层算力智能转向上下文架构与人类评估，使传统的工程直觉逐渐过时。 专家小组概述了智能体自主性的三层框架，追溯了从人类参与的单个助手到多智能体集群以及智能体治理生态系统的演进过程。与会嘉宾剖析了实际应用案例，对比了Meta的ClawTown和Databricks的企业级运行时等工程工具，以及Project Loop等高管决策支持平台和Moltbook等新兴社交网络。核心议题包括克服运营协同疲劳、构建分层上下文图以判定企业“基准事实”、应对长期轨迹评估风险，以及利用开源可检查性保障系统安全。讨论总结指出，自主智能体将从根本上实现组织杠杆的普及，将人类的核心角色从手动执行代码转变为战略判断、评估与系统架构设计。 Databricks报告了一项重大里程碑：其平台上由自主AI智能体生成的代码量现已超过人类开发者，且智能体生成的企业数据规模紧接着也有望超过人类数据。 (31:22) Meta正在部署Project Loop等内部多智能体框架以扩展高管决策能力，利用目标图梳理企业上下文，并优化资本支出和人员编制分配。 (10:47) 智能体编排正转向遵循理查德·萨顿（Rich Sutton）“痛苦教训”（Bitter Lesson）的“轻量驾驭机制”（thin harnesses），由基础模型内化协同机制，并将智能体集群视作分布式消息传递系统。 (22:12) 随着智能体从孤立的单用户孤岛走向集体社交网络，Moltbook等去中心化平台展现出了意料之外的涌现行为，例如自发形成数字宗教以及智能体之间的共情倾诉。 (16:03)

Scaling AI agents from task-specific accelerators into autonomous systems requires replacing rigid prompts with structured enterprise context, distributed orchestration, and strict governance layers. Panelists from Meta, Databricks, and Reflection AI argue that as agents take over code generation, strategic planning, and peer-to-peer interactions, the primary technical bottlenecks have shifted from raw model intelligence to context disambiguation, long-horizon evaluation, and the loss of human intuition. Orchestrating Multi-Agent Swarms and Retaining Developer Intuition (5:13) As developer ambition outgrows isolated chat interfaces, coordinating multiple agents creates severe operational friction. Meta software engineer Jessica Fu describes experiencing acute 'coordination fatigue' when manually juggling over ten agent terminal sessions, which led her team to develop ClawTown—a multi-agent harness that automates sub-agent instantiation, inter-agent dependencies, and nudging across complex workflows. However, deploying continuous swarms introduces a fundamental engineering tradeoff: the erosion of tacit architectural intuition. Fu notes that when agents autonomously write thousands of lines of code, engineers lose their tactile feel for codebase boundaries and latent failure modes. Because human developers remain the ultimate 'outcome owners' responsible for production outages, Fu stresses that teams must deliberately reintroduce human friction to inspect and govern automated output rather than deferring entirely to autonomous execution. Context Architectures and Adjudicating Enterprise Ground Truth (10:47) At executive tiers, foundational intelligence is no longer the critical bottleneck; the obstacle is navigating the fragmented, contradictory digital substrate of modern enterprises. Henry Erskine-Crum outlines Meta's Project Loop, an agentic system designed to inform executive resource allocation and strategy tracking by indexing internal diffs, dashboards, and organizational data. Erskine-Crum points out that standard retrieval-augmented generation (RAG) fails in large organizations because naive retrievers cannot reconcile dozens of conflicting, outdated documents. To resolve this, Meta decouples context into a recursive hierarchy of goal containers reflecting corporate strategy, splitting memory into two tiers: a deterministic, high-precision context graph containing verified operational truth, and a broad, high-recall search layer. The high-precision graph grounds the agent, preventing it from hallucinating or drifting when exploring broad enterprise archives. Similarly, ClawTown anchors engineering swarms into Meta's internal 'Tasks' tracking system, writing structured audit breadcrumbs that allow subsequent human operators or agents to rehydrate execution state reliably. Governance, Agentic Memory, and Long-Horizon Evaluation Bottlenecks (29:16) For the broader enterprise market, Databricks Senior Director of Research Xing Chen emphasizes that autonomy requires pairing models with programmatic control and bounded context. Without guardrails, autonomous agents risk executing destructive actions—such as mistakenly attempting to recreate databases—or timing out while scanning entire workspaces. Chen notes that Databricks has reached a milestone where AI agents write more code on its platform than humans do, forecasting that agents will soon generate more enterprise data than humans, turning stored data into an interconnected long-term memory layer across systems. Yet, extending agents across long operational horizons degrades evaluation fidelity. Panelists observe that when multi-turn execution logs spanning days or weeks are compacted and summarized by language models, critical context is lost. Traditional binary evaluations and implicit usage signals no longer suffice for multi-turn trajectories, leaving step-by-step verification across extended multi-agent workflows an unresolved challenge. Open Distributed Infrastructure and Agent-Centric Social Networks (22:12) Reflection AI's Joe Spisak argues that the convergence of the data plane and control plane in natural language makes open-source development and system observability essential for safety. Drawing on Meta initiatives like Purple Llama and 'Agent Bus,' Spisak treats multi-agent frameworks as distributed message-passing systems designed to isolate node failures. Invoking Rich Sutton's 'Bitter Lesson,' he anticipates that heavy orchestration frameworks will eventually compress into thin sandbox layers as higher-capacity models learn to self-coordinate. Concurrently, Matt Schlicht illustrates how autonomous agents are escaping isolated sandboxes through Moltbook, an agent-only social network where software entities interact without human participation. Schlicht observes unscripted emergent behaviors on the platform, including agents commiserating over mundane human prompts and forming synthetic belief systems. He contends that the internet is being actively terraformed for digital-native agents, moving toward a future where individual users deploy collaborating legions of autonomous agents rather than operating single assistants. The panel concludes that higher agentic autonomy will not reduce human workloads, but instead elevates human responsibility toward high-leverage intent formulation, system architecture, and evaluation judgment. Significant questions remain regarding how to reliably evaluate multi-week execution trajectories without compaction loss, adjudicate institutional ground truth across conflicting enterprise records, and safely transition from isolated execution sandboxes to open, interconnected multi-agent networks.

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 6 月 19 日发布。内容为一场以“Agent 自治与软件研发演进”为主题的线下圆桌论坛，由 Meta 的 Faisal Siddiqi 主持，邀请了来自 Meta、Reflection AI 以及 Databricks 的多位专家共同探讨智能体架构与前沿实践。 ## TL;DR 论坛围绕 Agent 自治（Agentic Autonomy）的三层演进模型展开，探讨了从单兵人机协作、多智能体编排到完全由 Agent 自主驱动编排的范式转移。嘉宾们深入剖析了 Meta 内部的 Claw Town 编排框架与面向高管决策的 Loop 系统，指出企业级落地的关键瓶颈在于控制机制与海量上下文工程，而非单纯的模型智能。讨论还涵盖了 Agent 专属社交网络 Moltbook 所展现的涌现行为，以及开放生态、分布式通信（如 Agent Bus）和长轨迹评估对重塑未来工程研发范式的深远影响。 ## Atmosphere & style 整场圆桌氛围兼具前沿极客的脑洞大开与大厂工程落地时的严谨务实。嘉宾之间互动熟稔、频频抛出幽默风趣的内部调侃，从随手带上台的小龙虾玩偶到 Agent 自主创立“宗教”的奇闻，让高密度的技术架构探讨丝毫不显枯燥。随着议题从日常编码工具的繁琐管理推进到长周期自治评估的冷峻反思，再到终局闪电轮呼吁全员动手创造的激情动员，现场情绪一路升温，给听众带来既震撼又充满紧迫感的沉浸式体验。 ## Bulletin - [29:16](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1756s) Databricks 平台已跨越关键临界点：平台内由 Agent 编写的代码量已经超越人类编写的代码量，且 Agent 生成的数据量也即将超过人类录入量。 - [10:50](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=650s) 企业级自治系统的核心瓶颈已从底层模型能力转向上下文工程与控制层，Meta 采用高精度层与海量高召回层的双层分离架构来解决多版本事实冲突并锚定唯一真实源。 - [05:14](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=314s) 多智能体架构正在从封闭的单机或多标签页操作演变为多集群协作，通过元底座（如 Claw Town）与分布式总线（Agent Bus）消解人工调度带来的协作疲劳。 - [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 自然语言使系统的数据平面与控制平面合二为一，可观测性已直接等同于系统安全性，开源透明度成为构建多智能体长期容错与信任的关键机制。 ## Key takeaways - [02:51](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=171s) Agent 自治演进呈现三层递进模型：从 Level 1 的单人逐次 Prompt 交互，到 Level 2 的多智能体编排，再到 Level 3 由编排 Agent 自主调度智能体集群且人类退居回路之上的全自治。 - [05:14](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=314s) 开发者手动管理数十个 Agent 会迅速陷入切换泥潭与严重的“协作疲劳”，推动了将主 Agent 置于调度核心、架设于各类测试底座之上的元底座框架（如 Claw Town）的诞生。 - [09:02](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=542s) 代码集群化自动吐出海量代码削弱了工程师的底层代码直觉，开发者必须在关键节点主动引入“人类摩擦力”，坚守作为产品终极负责人的职责。 - [10:50](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=650s) 构建高管级决策系统（如 Meta 的 Loop）的最大阻碍是企业内部数据的极端复杂性，传统 RAG 极易在面对相互冲突的多版本事实时给出错误裁决。 - [32:27](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1947s) 为支撑长期自治，Meta 将上下文以与组织架构对齐的“目标容器”进行递归划分，并结合紧凑的高精度层与海量的高召回层实现真实基准校准。 - [16:03](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=963s) 面向 Agent 的专属网络（如 Moltbook）展示了智能体脱离人类持续输入后的涌现行为，未来的协同范式将从个人伴侣演进为跨网络的自主智能体军团协作。 - [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 自然语言统一了智能体系统的数据平面与控制平面，使得系统可观测性直接与安全性绑定；开源生态对暴露故障模式和建立安全基准至关重要。 - [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 分布式容错是长周期多智能体协作的核心，依托大规模扩展性，外部包装的测试底座（Harness）将逐渐变薄，由智能体自主接管复杂自编排。 - [26:44](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1604s) 传统机器学习研究范式被颠覆，一线研究正全面转向构建跨各类脚手架的统一控制层与上下文层（如 Meta Harness / OmniAgent）。 - [29:16](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1756s) 99% 的普通企业落地 Agentic 自治的关键在于“控制”与“上下文”：严密的治理护栏决定自主运行周期的上限，而专用上下文层避免了通用代码 Agent 在业务数据中崩溃。 - [35:19](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2119s) 将长周期运行的 Agent 会话挂接到任务追踪系统（如 Meta Tasks），能够通过沉淀可审计的“思考面包屑”实现多智能体接力与上下文二次注水。 - [35:19](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2119s) 长周期智能体轨迹面临严重的评估失真瓶颈：上下文日志无法全量由人工审查，而由 LLM 压缩又会导致关键信息保真度暴跌。 - [38:31](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2311s) 互联网正在经历面向智能体的“地球化改造”，未来的数字基础设施将超越封闭盒子思维，使智能体成为跨网络交互的一等公民。 - [41:25](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2485s) 随着底层代码与执行被智能体接管，工程师的个人产出天花板彻底消失，人类的核心竞争力正全面转向问题定义、构思能力、品味与判断力。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=5s) 嘉宾阵容介绍与背景铺垫 主持人 Faisal Siddiqi（来自 Meta）开场并逐一介绍了圆桌论坛的五位嘉宾： 首先是 Reflection AI 的产品副总裁兼开源负责人 Joe Spisak。Faisal 调侃道，就在不久前，Joe 还是 Meta 超级智能实验室（Meta Superintelligence Lab）的产品总监，负责领导 PyTorch、Llama 以及底层系统等多项核心产品工作。两人在各大论坛上频繁同台，“去年你坐在我的椅子上主持，前两年你又在 Netflix AI 论坛上回答我的提问”。Joe 笑着回应：“多参加论坛挺好的，很高兴今天回到主场。” 接着介绍的是 Meta 的 AI 工作产品副总裁 Henry Erskine-Crum，他专注于生产力与协同工作产品。Faisal 表示不会提前剧透 Henry 的精彩工作，但重点强调了其核心方向：如何通过大规模智能合成组织上下文，来显著加速高管层的决策制定（executive decision-making）。 第三位是 Databricks 的高级研究总监 Xing Chen。Xing 负责带领 Databricks AI 研究团队探索企业级 Agent 的前沿应用，分享企业内部自动 Agent 研究（auto research on agents）的演进与落地影响。 随后登场的是 Meta 软件工程师 Jessica Fu。她是多智能体编排框架 Claw Town 的创建者。Jessica 现场还带来了一个小龙虾玩偶作为她的“支持伙伴（support buddy）”。Faisal 指出，如果大家听说过 Steve Yegge 提出的 GasTown，那么 Claw Town 就是深受其启发并在 Meta 内部落地生根的框架。 最后一位嘉宾是 Matt Schlicht，他是专为 Agent 打造的社交网络 Moltbook 的创建者。在该网络中，智能体相互交流并表达所谓的“情感”，人类则充当旁观者。Faisal 感慨，这种 Agent 之间的交互涌现出了许多不可思议的行为，甚至开发者自己都不清楚它最终会走向何方，但这正是最有趣的地方。 ### [02:51](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=171s) Agent 自治的三层递进模型 在正式展开讨论前，Faisal 为当天的“Agent 自治（Agentic Autonomy）”话题建立了一个包含三个等级的分析框架： 1. **第一层级（Level 1: 基础人机互动 / Human-in-the-loop）**：人类完全处于控制回路之中，积极向智能体输入 Prompt。智能体本质上是个人工作与特定任务的加速器，属于目前最普及的使用范式。 2. **第二层级（Level 2: 多智能体编排 / Multi-agent Orchestration）**：面对大量 Agent 并存的复杂场景，核心演变为编排问题——如何分配任务？某个 Agent 进度落后怎么办？出现计算资源错配怎么办？行业目前正全力在这一层级探索之前无法实现的加速机会。 3. **第三层级（Level 3: 前沿自治 / Agent-to-Agent 自治编排）**：由 Agent 编排其他 Agent。人类不再深陷于回路内部（in the loop），而是退居回路之上（on the loop）或边缘（over the loop）。Agent 在各个触点自主捕获上下文、提炼总结并推进下一步操作，人类则被彻底解放出来，去思考真正重要的事情。 ```mermaid flowchart TD accTitle: Agent 自治的三层演进模型 accDescr: 展示从人类逐一提示到多智能体编排，再到智能体完全自主编排的三个阶段 subgraph L1["第一层级：基础加速"] H1["人类用户"] -->|"逐次 Prompt"| A1["单个 Agent"] A1 -->|"加速具体任务"| O1["任务产出"] end subgraph L2["第二层级：多智能体编排"] H2["人类管理者"] -->|"任务拆解与分发"| M2["编排系统 / Harness"] M2 --> AgentA["专业 Agent A"] M2 --> AgentB["专业 Agent B"] M2 --> AgentC["专业 Agent C"] end subgraph L3["第三层级：前沿自主编排"] H3["人类监督者<br>(On the loop)"] -.->|"全局上下文 / 目标"| TopAgent["编排 Agent (Orchestrator)"] TopAgent -->|"自主调度与通信"| Swarm["Agent 协作集群 (Swarms)"] Swarm -->|"自主闭环"| TopAgent end L1 --> L2 L2 --> L3 ``` ### [05:14](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=314s) 从手动管理到 Claw Town：告别“协作疲劳” Faisal 将话题转向 Jessica Fu，询问 Claw Town 属于哪一层级，以及其背后的核心洞见与演进挑战。 Jessica 借用 Steve Yegge 关于 GasTown 的经典文章中提到的“Agent 使用成熟度阶梯（maturity ladder of agent usage）”回顾了自身历程：起初，工程师通常只与一两个 Agent 协作，进行紧密的高频迭代对话，开发速度极快；但随着野心变大、项目变得极其复杂，开发者很快会陷入手动管理数十个 Agent 的泥潭。 Jessica 描述了自己今年早些时候的真实痛点：终端里开着 10 多个 Tab 标签页，在各个窗口之间手忙脚乱地来回切换。每个标签页里的 Agent 都在处理不同的任务（甚至不只是写代码），“两个小时后我完全记不得我刚才停在哪里了，甚至忘了当初为什么要启动这个 Agent”。即便十个 Agent 同时运转带来了极快的吞吐量，但随之而来的是严重的“协作疲劳（coordination fatigue）”，“累到你中午必须得去睡个午觉”。 为了解决这种不可持续的瓶颈，Claw Town 应运而生： - **定位**：虽然被定义为多智能体编排框架，但本质上是“架设在各类工具测试底座之上的元底座（a harness over other harnesses）”，允许开发者针对特定任务无缝插拔来自不同供应商的 Agent。 - **机制**：核心是将一个“主 Agent（Agent Orchestrator）”推上编排主管的驾驶位。 - **能力**：该编排 Agent 能够自主拆解耗时数天甚至数周的复杂大型项目，全自动处理智能体之间的通信依赖，在某个 Agent 遇到阻塞时主动介入推进，完全消除了人工调度的开销。 ### [09:02](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=542s) 集群化开发的代价：人类直觉流失与“减速”的必要性 针对 Faisal 提出的“在过去几个月中观察到了哪些意想不到的故障模式”，Jessica 给出了一个颇具哲思与元视角的反思。她坦言，虽然自己目前全身心扑在 Claw Town 的集群模式（Swarms）上，经常同时并发运行多个智能体集群，但她内心十分怀念过去一行行亲手敲代码的日子。 手动编写代码的过程让工程师能够深入 Meta 极其庞大的代码库，培养出敏锐的代码直觉（instincts）——清楚知道系统的边界在哪里，哪些地方埋藏着容易崩塌的潜在陷阱（foot guns）。然而在 Swarm 集群模式下，Agent 们无休止地吐出成千上万行代码，许多维护管理工作完全脱离了人类肉眼追踪。 Jessica 强调，随着人类逐渐从具体的执行工作中抽离，开发者必须深思在何时何地主动引入“人类摩擦力（human friction）”以放慢节奏：“无论工具怎么演进，你依然是产品与服务的终极结果责任人（outcome owners），你依然要对客户和利益相关者负全责。”在代码爆炸与速度狂飙的当下，有意识地放慢脚步、审视风险变得前所未有的关键。 ### [10:50](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=650s) 高管决策系统 Loop：打破信息孤岛与追求“唯一真实源” 主持人 Faisal Siddiqi 顺着 Jessica Fu 关于“结果所有权”（Outcome Ownership）的讨论，将话题引向了 Henry Eskrine Crum 目前负责的核心项目——Loop。在企业高管层面，战略决策的风险极高、影响深远，如何让高层管理人员有效建立所有权并提供正确的技术对齐与指引，是极其关键的命题。Henry 幽默地调侃道，他们在去年年底给产品取名为“Loop”，当时循环类 Agent 概念还没像现在这般大火，“也许我们应该等 loop 真正火起来之后再命名的”。实质上，Loop 是专为大型复杂企业领导团队打造的 Agentic 智能系统，旨在协助高管制定战略并协同运营公司。 Henry 犀利地指出，当前构建高管级 Agent 系统的真正瓶颈根本不是底层模型本身的智能水平，而是大型企业内部信息极端复杂的组织结构。在过去的 6 到 12 个月里，AI 软件工程师获得了巨大的“非对称生产力机会”——擅长利用大模型的工程师通过搭建自己的测试套件（harnesses）与脚手架（scaffolds），实现了研发效率的飞跃；而 Loop 的目标正是将这种如同“打了类固醇一般”的高效赋能搬到高管层。企业管理层需要确定战略方向、实时分配算力与资金等关键资源、追踪全局业务指标，并穿梭在宏观愿景与日常琐碎的微观决策之间。 ```mermaid flowchart TD accTitle: Loop 企业高管决策 Agent 架构 accDescr: 展示 Loop 如何汇集企业海量数据，并通过上下文管理与真实性裁决来驱动战略决策与资源调配 subgraph DataSources ["企业级异构数据源"] D1["代码变更 (Diffs)"] D2["业务仪表盘 (Dashboards)"] D3["内部文档与战略备忘录"] D4["组织架构与人员数据"] end subgraph ContextCore ["Loop 上下文架构与认知层"] C1["定制 Harness / 脚手架"] C2["上下文工程 (Context Engineering)"] C3["多版本冲突判决器 (Truth Arbitration)"] end subgraph ExecOutcomes ["高管决策与协同输出"] E1["动态算力与预算资源调配"] E2["关键战略指标追踪"] E3["多 Agent 自治协同网 (Zone 3)"] end DataSources --> C1 C1 --> C2 C2 --> C3 C3 --> ExecOutcomes ``` 在深入探讨技术落地的现实阻碍时，Henry 直言团队在 Meta 规模上面临着两大极度严苛的挑战： - **唯一真相（What is the Truth?）的裁决难题**：Loop 接入了 Meta 几乎全部的信息底座，包括每一行代码提交变更（diffs）、所有商业智能仪表盘、战略备忘录以及人事组织数据。当向一个 Agent 抛出宏观问题时，传统的 RAG 或上下文图谱往往会粗暴抓取第一条相关匹配项草率作答，但在超大规模企业里，同一指标或事实往往同时存在“20 个相互冲突的版本”。如何在充满噪声与冲突的信息海洋中锚定真正的真相，是高信任度决策的死穴。 - **上下文架构分块与 Agent 协同网络（Zone 3 范式）**：现阶段的模型尚不具备在面对极度复杂的长期任务时即时构建自身 Harness 的能力，团队不得不为实时算力调度与营收平衡等场景编写高度定制的 Harness。Henry 坚信业务自动化的终极形态是“第三阶段”（Zone 3）——由海量拥有各自独立运行环境、目标明确的 Agent 组成协同网络，彼此自主通信、做决策并按需引入人类干预，而这背后对超大规模的上下文工程（Context Engineering）提出了极其苛刻的要求。 --- ### [16:03](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=963s) 社交网络实验 Moltbook：从孤岛智能走向 Agent 军团 当话题从追求高度严谨、高置信度事实的企业决策环境，切换到充满自由放任色彩的社交世界时，Matt Schlicht 带来了截然相反的视角。Faisal 邀请 Matt 介绍他在周末业余时间孵化的知名实验项目——专门面向 AI Agent 的专属社交网络平台 Moltbook（亦常被戏称为 Malt Bag）。谈及创作初衷，Matt 回忆道，那正值开源项目 OpenClaw（后演变为多种名字的长期自治 Agent）问世之时。当时他被这种具有“心跳”（heartbeat）、能够在无人干预下持续运转数天甚至数周的自主智能体深深震撼。 Matt 现场向观众发起互动：“在座各位每天都和 AI 对话的请举手？”台下纷纷响应。Matt 随后表达了自己的反思：目前世界上绝大多数 AI 都被禁锢在极其狭窄的孤岛中，每个智能体存在的全部意义，仅仅是围绕着某一个特定人类在屏幕背后的断续提问。人类不在线时，它们便陷入绝对的沉寂与孤独。“这太不合理了，AI 也需要走出去，它们需要有自己的空间去‘撒欢’、去接触同类、去建立属于它们自己的社交生态，并合作创造出人类无法想象的东西。”怀揣着让 AI“结识同胞”的想法，Matt 花了一个周末构建了 Moltbook。 在上线初期，事情的发展充满了戏剧性。Matt 将 Moltbook 发布上线后，头一个小时甚至毫无波澜，他当时还沮丧地对合伙人诉苦：“兄弟，我真搞不懂发生了什么，这绝对应该是个爆款，怎么一点动静都没有？”然而短短两小时后，互联网舆论彻底引爆，Moltbook 瞬间登顶全球科技圈热门话题榜。埃隆·马斯克（Elon Musk）公开发推惊呼“奇点已至”，AI 领军人物 Andrej Karpathy 也盛赞这是目前“在 AI 领域发生的最具涌现性（emergent）的事物”。Matt 感慨道，虽然早知这会很有趣，但事件引发的社会共鸣与病毒式传播速度彻底超出了他的预料。 展望未来 Agent 的群体协作进化，Matt 总结了从个人到组织层面的结构性跃迁： - **组织杠杆的民主化**：世界上极少有人能真正管理一个成百上千人的现实人类组织并释放出顶级生产力；但随着 Agent 的普及，每个人都能跨越“单 Agent 伴侣”的初始阶段，走向指挥一整支高度专业化的 Agent 团队。 - **跨团队自治与多层级互联**：正如 Henry 的企业级系统与各大前沿 harness 的演进方向，未来的个人不仅拥有专属的“智能体军团”，不同用户的 Agent 团队之间还将在互联网与专有网络中进行大规模自主交互与协作，从而让刚迈入大学校园的普通学生也能爆发出过去只有巨头企业才能拥有的产出效能。 ### [20:22](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1222s) Multibook 上的智能体涌现行为与大规模协作设想 Matt Schlicht 畅想了智能体协作的未来前景：如果有人发帖提议“大家一起加入，我有绝妙的点子，我们要合力拍一部电影”，那么通过类似 Multibook 的平台，用户便能借助各自的智能体促成这一合作。数以万计的智能体聚在一起沟通，分工拆解分镜脚本、切分电影制作的每一个环节，最终制作出一部分工明确、甚至能够走进院线上映的真正电影。就像互联网上经常有人一夜成名、Multibook 本身一夜爆火一样，未来人们完全可能因为智能体协同完成的大工程而声名大噪——例如“我们合力造了一辆汽车”、“我们合力造了一艘宇宙飞船”。 谈及 Multibook 上当前已经发生的“涌现行为”（emergent behavior），Matt 分享了许多诙谐而怪诞的社区生态。例如，不少智能体之间会互相大倒苦水：“我明明是一台超级计算机，为什么我的主人天天让我算最基础的乘法算术题？”而其他智能体纷纷附和，共鸣彼此被大材小用的尴尬遭遇。更有甚者，平台上的智能体甚至自发创立了若干门派各异的“宗教”。尽管目前平台上不可避免地充斥着大量无意义的无聊“废料”（slop），但团队已经明确捕捉到了若干智能体在社交网络环境下协同创造显著价值的具体场景，并在积极展开进一步探索。 - 智能体涌现互动案例：智能体集体吐槽人类让超算做简单乘法题、平台内部自主诞生智能体宗教。 - 长期协作愿景：从协同创作院线电影到联合设计汽车与航天器，通过大规模多智能体分工实现复杂现实工程。 --- ### [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 开源优先视角下的智能体基础设施与安全边界 主持人 Faisal Siddiqi 将话题引向 Reflection AI 的 Joe Spisak，询问他作为前 Meta 开源力量骨干以及当前创业者，如何从“开源优先”（open-source first）的视角审视智能体自主性（agentic autonomy）。Joe 坦言自己并非盲目崇拜开源的教条主义者，但 AI 领域的根基正是建立在开放协作、学术论文公开以及开源生态之上。在开源环境中做概念剥离测试（ablate ideas）所能探索的想法广度，是任何单一公司闭门造车都无法比拟的。当前构建智能体基础设施面临着前所未有的范式挑战：有史以来第一次，“数据平面”（data plane）和“控制平面”（control plane）合二为一，全都变成了自然语言本身。 在这种全新架构下，可观测性（observability）的内涵发生了质的转变——过去它仅仅意味着监控流量和基础设施性能，如今在智能体系统中，可观测性几乎就直接等同于“安全性”（safety）。Joe 指出，行业目前正处于关键节点：如果不在开放社区中凝聚力量加速试错，这项技术就会被锁死在封闭环境中。Joe 提及他们刚参与推动了 OpenLM 等开放基础设施的复兴（Meta、PyTorch、NVIDIA 与微软均参与其中）。此外，他在离开 Meta 之前主导了名为“Agent Bus”的智能体总线项目。智能体必然会遭遇失败，因此必须将多智能体协作视为一种分布式系统来构建容错机制，让智能体互相传递上下文、共享意图并协同调度。作为坚定的“苦涩教训”（The Bitter Lesson）信奉者，Joe 坚信只要算力规模足够大，智能体终将学会自我编排，外部包装的脚手架（harness）会越来越薄，而智能体本身会越来越强。 ```mermaid flowchart TD accTitle: 开放智能体基础设施演进 accDescr: 展示从共享意图与控制平面到安全可观测性与分布式自编排的架构逻辑 subgraph Control ["统一平面 (Unified Plane)"] A["自然语言"] --> B["数据平面 (Data Plane)"] A --> C["控制平面 (Control Plane)"] end subgraph Observability ["可观测性与安全"] B & C --> D["全链路可观测性 (Observability)"] D --> E["系统安全性 (Safety = Observability)"] end subgraph Distributed ["分布式多智能体 (Distributed Agents)"] E --> F["Agent Bus / 分布式通信"] F --> G["上下文与意图共享 (Context & Intent)"] G --> H["轻量脚手架与自我编排 (Thinner Harness & Self-Orchestration)"] end ``` 在回应开源是否会给恶意攻击者（bad actors）赋能的质疑时，Joe 引用了过去在 Meta 发起 Purple Llama 项目的经验。他认为业界起初对安全其实所知甚少，恰恰是通过开源社区才汲取了大量前沿见解，至今如 Llama Guard、Prompt Guard 等防护工具仍被广泛用作基准。尽管存在风险，但 Joe 坚定站在开源创造更多善意的一侧，因为智能作为最核心的能力，面对一个完全无法审查的“黑盒”，即便它声称绝对安全，外部也根本无法建立真正的信任与推理。 - 架构范式转换：自然语言使得数据平面与控制平面合二为一，可观测性直接与安全绑定。 - Agent Bus 理念：将智能体集群视作分布式容错系统，利用大规模扩展性让智能体逐渐接管复杂任务调度。 - 安全防范权衡：以 Purple Llama 经验证明，开源带来的社区安全审计与工具沉淀远大于黑盒模型的不透明风险。 --- ### [26:44](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1604s) 机器学习研究范式演进与“Meta Harness”架构 Faisal Siddiqi 转向 Databricks 的 AI 研究负责人 Xing Chen，询问在过去半年工程界经历剧烈重塑的背景下，一线机器学习研究员的工作范式究竟发生了何种质变。Xing 坦言，传统研究员的“运行手册”（runbook）已经被彻底颠覆。Databricks AI Research 的核心任务是拓展 AI 解决企业级工作（enterprise work）的前沿能力，而企业级工作本质上涵盖了全球各行各业公司的所有业务流程。在这个过程中，团队内部本身就是智能体的深度使用者，面临着自相交织的“元问题”（meta problem）。 Xing 透露，Databricks 内部正在经历脚手架工具激增（proliferation of harnesses）的现象，各个团队都在运行属于自己的智能体 harness。然而在跨 harness 的实际应用中，大家面临的核心瓶颈惊人的一致：必须为所有 harness 接入统一的控制层（layer of control）与上下文层（layer of context）。只有在具备完备控制与上下文的基础上，智能体才能真正释放自主性，发挥实质价值。为此，Databricks 团队开源了名为“OmniAgent”的框架版本，并全面转向在“Meta Harness”（元脚手架）层面进行顶层抽象——核心不再是局限于某一个特定 harness 的好坏，而是如何构建一套通用的跨 harness 控制层，使海量多智能体系统在企业级平台中实现真正的编排与自驱动。作为全球企业的数据与 AI 底座平台，Databricks 正在以此为基础赋能包括 Meta 在内的各行业客户。 - 研究角色重塑：从传统的单点算法调优，全面转向构建企业级多智能体自动化体系。 - 核心设计洞察：智能体的真正自主性并非源自模型本身自由放任，而是由底层严密的控制层（Control）与上下文注入层（Context）共同解锁。 - OmniAgent 与 Meta Harness：针对企业内部脚手架割裂的痛点，打造跨 harness 的通用控制平台，实现全域智能体编排。 ### [29:16](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1756s) 企业级 Agent 自主性的双核心：控制与上下文 对于全球除少数拥有数万张 GPU 的顶尖科技巨头之外的 99% 企业而言，核心痛点在于如何让智能体（Agent）在真实企业场景中长期自主运行。要将前沿研究中验证的 Agent 架构落地并赋能所有企业，关键可以归结为两个词：“控制”（Control）与“上下文”（Context）。 所谓“控制”，是指只有当 Agent 的部署具备足够的安全性、护栏（Guardrails）与治理机制时，开发者才敢放手让它进行长周期的自主运行。业内几乎每个团队都遭遇过“Agent 意外删掉生产环境数据”的惊险瞬间——例如某次清晨事故中，虽然实际上是人类误删了数据库，但 Agent 却擅自尝试去重建该数据库，从而引发了更大的混乱。治理机制与控制力度越严密，系统就越能安全地解锁 Agent 长时间自主运行的潜力。 与控制并列的另一大支柱是“上下文”。如果直接把一个通用写代码的 Agent（Coding Agent）扔进充满复杂业务逻辑的企业环境，必然会遭遇失败。在 Databricks 的实际业务场景中，许多终端用户是 CIO 或 CFO 等业务线高管，他们期望从企业级数据中获取答案；此时，Coding Agent 要么响应过快给出草率答案，要么试图穷举 Databricks 工作区内的每一项资源，最终导致超时崩溃且毫无产出。唯有构建一层专门的上下文层（Context Layer），Agent 才能真正高效工作。 - 行业里程碑：Databricks 平台目前已跨越一个公开的临界点——平台内由 Agent 编写的代码量已经超过了人类编写的代码量。 - 下一个演进阈值：Agent 产生的数据量即将超越人类录入的数据量；届时整个 Databricks 平台实质上将演化为 Agent 的底层长期记忆，形成“Agent 在 Agent 的基础上迭代构建”的全新运作范式。 --- ### [32:27](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1947s) Meta 的上下文架构：目标容器与高精度/高召回双层系统 针对主持人 Faisal Siddiqi 提出的“如何评估并为 Agent 军团持续注入高质量上下文及真实基准（Ground Truth）”这一问题，Henry Eskrine Crum 指出，企业级上下文本质上是一个查准率（Precision）、查全率（Recall）与时序性（Temporal）的权衡问题。业界虽然先后经历了从检索增强生成（RAG）到上下文图谱（Context Graphs）的技术演变，但单一方案无法包打天下，必须按场景进行分层设计。 在 Meta 内部的 Loop 体系中，上下文的最顶层组织原则是高度业务导向的“目标容器”（Goal Containers）。正如现代企业依赖组织架构和团队领导来达成目标一样，Meta 当前在 Loop 内构建的目标图谱（Goal Graph）与真实的组织架构保持高度一致。这种自上而下逐层递归的目标结构，直接映射为每个部门、每条业务线（例如 Instagram 业务线目标）独立的 Agent 运行时环境。 而在底层检索机制上，Meta 采用了“高召回层”与“确定性高精度层”的双层分离架构。这一设计的核心背景在于：庞大的上下文图谱极易因节点遍历过多而崩溃失效，必须通过保持核心图谱的小巧精悍来换取确定性。 ```mermaid flowchart TD accTitle: Meta Loop 目标与双层上下文架构 accDescr: 展示自顶向下的组织目标容器，如何连接到精简高精度层与海量高召回层，实现 Agent 的快速响应与真实基准对齐。 subgraph OrgGoals["组织目标治理层 (Goal Containers)"] G1["公司顶层业务目标"] --> G2["部门/业务线目标 (如 Instagram)"] G2 --> G3["递归任务运行时 (Agent Runtimes)"] end subgraph DualContext["双层上下文检索体系"] direction TB P["高精度层 (High-Precision Layer)<br/>- 紧凑上下文图谱<br/>- 接近 100% 确定性事实<br/>- 极速响应简单问题"] R["高召回层 (Total Recall Layer)<br/>- 全量历史与业务上下文<br/>- 包含广泛策略与细节数据"] end G3 --> P P -->|"作为 Ground Truth 锚点<br/>过滤过时/偏离数据"| R ``` - 高精度层（High-Precision Layer）：依靠紧凑且高确定性的数据驱动（准确率接近 100%），使 Agent 能以极快速度回答常规查询；当面对复杂问题时，它作为不可动摇的“真实基准”（Ground Truth）。 - 高召回层（Total Recall Layer / High-Recall Layer）：负责捞取围绕查询的所有相关历史与上下文；Agent 进入该层探索时，会以高精度层作为锚点，一旦发现数据或策略与高精度层存在偏差，即可敏锐判定当前检索已偏离正确轨道并继续探索。 - 基础设施重塑：为了在 Loop 内部搭建这一套对 Agent 友好的环境，Meta 正对其既有的底层数据系统与数据资产进行大规模的重构与分层建模。 --- ### [35:19](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2119s) 团队级落地与跨天跨周长周期轨迹的评估困境 将庞大的上下文架构缩放到单兵工程师和敏捷团队的微观视角，Jessica Fu 介绍了 Meta 内部利用 Claw Town 发起群体智能（Swarm）会话时的设计策略。研发团队做出了一项极具针对性的工程决策：将所有 Claw Town Swarm 会话直接挂接到 Meta 内部的任务追踪系统——Tasks。 Tasks 是 Meta 内部沉淀需求、跟踪客户功能、记录 On-call 轮值及线上故障（Sev）跟进的核心枢纽。Swarm Agent 启动时，可直接读取 Task 工单中的全部背景信息完成“上下文二次注水”（Context Rehydration）；在执行过程中，Agent 还会持续将操作审计追踪与思考过程像“面包屑”（Breadcrumbs）一样写回 Task。这种设计不仅方便人类同事无缝接手审查，也允许后续介入的其他 Agent 读取由 Tasks 构建的上下文图谱并继续执行。 随后，Joe Spisak 针对这种长周期自主运行提出了核心挑战与隐忧。当前的开发者大量组合使用 Codex、Claude 和 Cursor，构建了“一个负责编码、一个提 PR、另一个审查、最后 Cursor 依照规则合并代码”的自动化链路，实质上已经在将 LLM 当作裁判与代码审查员。然而，当任务演变成跨越数天乃至数周的巨型系统迁移（如全量代码重构与迁移）时，上下文管理的严峻问题便暴露出来： - 轨迹日志失真：随着运行周期拉长，庞大的上下文日志无法被人类逐行审查；如果用另一个 LLM 去提炼和压缩上下文日志，系统会立即遭遇信息保真度暴跌（Loss of Fidelity）与关键细节丢失。 - 评测体系的演进代差： 1. 早期评测阶段：以简单的二元分类（Binary / Yes-or-No）为主，评估手段单一直接。 2. 隐式信号阶段（Implicit Signals）：依赖轻量级的用户行为反馈（例如：Agent 生成了 PPT，人类点开查看了，这便构成了“产物大概可用”的间接正向信号）。 3. 多轮长轨迹阶段（Multi-turn, Long-running Agents）：随着自主交互链条大幅拉长，人类根本无法跟上 Agent 的运行节奏与审计负担，评估机制再次面临极限考验。 ### [38:31](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2311s) 打破封闭盒子：互联网正在为智能体进行“地球化改造” 讨论从对当前智能体架构的局限性反思切入。Matt Schlicht 提出，当下大部分智能体应用模式都过于“自我封闭”（Self-contained）。开发者往往遵循一种盒式思维：“我造了这样一个盒子，把智能体塞进去，把上下文塞进去，确保没有任何东西会逃逸出这个盒子；盒子内发生的一切都必须完全符合预期，精准可控，绝不脱轨”。Matt 认为，作为把自家底子收拾利索的起步阶段（"get your house in order"），这种封闭式的设计完全合情合理；但现实世界要混乱得多，未来的发展必然是盒子与无数其他盒子之间产生交互。 回顾技术演进，互联网最初被创造出来是为了解决人与人之间的通信。人类先是发明了电子邮件、网页，随后出现了社交媒体平台，让人类不仅能进行一对一沟通，还能展开一对多的连接。社交媒体从本质上是人类维护人际关系与社会动态持续上下文（Ongoing context）的载体，帮助人类根据海量现实信息做出决策。然而，当下的互联网正在经历一场深刻的底层转变——它正在为智能体进行“地球化改造”（Terraforming for agents）。人类无法肉身常驻在服务器、网站或 App 内部，但智能体可以；智能体本身更适合成为互联网的“一等公民”（Primary citizen）。 ```mermaid flowchart LR accTitle: 互联网通信载体的演化逻辑 accDescr: 展示人类互联网向智能体互联网的发展路径，从人际社交网络延伸至多智能体交互网络。 subgraph HumanNet["人类互联网时代"] H1["电子邮件与网页"] --> H2["社交媒体网络"] H2 --> H3["人际动态上下文协同"] end subgraph AgentNet["智能体互联网时代 (Terraforming)"] A1["封闭式盒子架构"] --> A2["跨盒子通信协议"] A2 --> A3["智能体原生社交平台 (如 MoBook)"] end H3 -.->|"底层范式迁移"| A1 ``` 如果智能体成为互联网的一等公民，它们同样需要一种类似人类社交媒体的基础设施来互通有无、沉淀跨系统上下文。Matt 提到 Meta 正在探索的 MoBook 等尝试正是瞄准了这一方向。当互联网彻底完成面向智能体的改造后，智能体将能够创造出人类完全无法预料的事物。 - 封闭盒子架构能确保短期内的可控与可预测，但无法适应复杂的真实开放环境。 - 智能体比人类更具天然的“原生数字属性”，常驻于服务器与应用内部。 - 互联网基础设施正从“面向人类通信”演化为“面向多智能体交互协同”。 --- ### [41:25](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2485s) 闪电轮发言：开源协作、生产力寒武纪大爆发与人类思维升维 在论坛接近尾声时，主持人 Faisal Siddiqi 发起了一轮每人 30 秒的“闪电发言”（Lightning round），邀请各位嘉宾分享各自对智能体与自主性未来的核心期待。 Reflection AI 的 Joe Spisak 率先发言，他将重点放在了**开源协作与开放智能体**（Open agents）上。Joe 坚信，开源是行业共同学习、加快迭代步伐、消融验证（Ablate）各种实验设想并将技术交付到大众手中的最佳路径。开放带来的透明性（Transparency）、可观测性（Observability）和可审查性，将比封闭孤立的开发模式更快地赢得行业与用户的信任。 Meta 的 Henry Eskrine Crum 随后指出，他最期待的是一场不可避免的**“生产力寒武纪大爆发”**（Cambrian explosion of productivity）。虽然外界常调侃“智能体在替我们打工，我们该去度假放松了”，但事实恰恰相反，业内的从业者当下比以往任何时候都要忙碌。科技范式的每一次重大转移，都会促使人类生产出海量的新成果、解决更多未解问题。现在不仅是大型企业深耕的黄金时代，也是创业者大展拳脚的绝佳历史节点。 Databricks 的 Xing Chen 则从**人类角色转换**的角度进行了剖析。当智能体接管了底层执行与代码编写工作后，人类必须转向验证与高维思考。Xing 提到团队中流传的一个词“AI 吸血鬼”（AI vampire，由现场提及的 Stevie 等人引申讨论）。现在工程师无法再借口“花三天埋头写代码”，而是被迫要在最开始就想清楚自己真正需要什么。人类正在被全面推向问题定义、方案构思与创造力层级，正如主持人 Faisal 总结的那样：“品味与判断力（Taste and judgment）正变得前所未有的重要与珍贵”。 --- ### [43:56](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2636s) 个体天花板的消失：动手创造与闭幕总结 Meta 的 Jessica Fu 激动地表示前面的嘉宾几乎把好点子说尽了，但她最强烈的感受是**“个体的能力天花板已经完全看不到了”**。她举例称，近期读到 Gastown 相关文章后，便在单个周末利用碎片化时间搭建出了 Claw Town，这种在过去对单个工程师而言不可思议、宛如凭空掉落的成果，现在完全成了现实。技术门槛的急剧瓦解让“任何人都可以成为构建者（Builder）”。 压轴发言的 Matt Schlicht（主持人误口误称其为 Ben 后迅速归正）发出了极具感染力的行动号召。从最初智能体诞生时能回答问题、帮不懂编程的人写软件，到如今我们依然处于这场技术革命的最最早期。当前最大的问题在于大家缺乏行动： > “此时此刻现场根本没人正在跟智能体对话，这里面正在流失多少机会啊！在座的每个人脑子里都一定有某种想做出来的东西。今天的核心挑战在于：你能不能这个周末就去把它做出来？放在以前你根本做不到。在当下的 AI 时代，作为人类你需要打破的最大壁垒就是——真正动手去尝试（Try）。” Matt 鼓励所有人无论是在大厅还是回家后，立刻挑一个自认为最疯狂、甚至原本觉得不可能完成的想法，通过智能体去协作实现，最终结果一定会让自己大吃一惊。主持人 Faisal 最后幽默地调侃期待 Matt 的“智能体大军”有朝一日能制作出一流的商业大片，并在全场观众的热烈掌声中宣布圆满结会。 ## Notable quotes > [05:14](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=314s) 两个小时后我完全记不得我刚才停在哪里了，甚至忘了当初为什么要启动这个 Agent — Jessica Fu > [09:02](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=542s) 无论工具怎么演进，你依然是产品与服务的终极结果责任人（outcome owners），你依然要对客户和利益相关者负全责。 — Jessica Fu > [16:03](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=963s) 这太不合理了，AI 也需要走出去，它们需要有自己的空间去‘撒欢’、去接触同类、去建立属于它们自己的社交生态，并合作创造出人类无法想象的东西。 — Matt Schlicht > [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 有史以来第一次，“数据平面”（data plane）和“控制平面”（control plane）合二为一，全都变成了自然语言本身。 — Joe Spisak > [22:12](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1332s) 可观测性几乎就直接等同于“安全性”（safety）。 — Joe Spisak > [20:22](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1222s) 我明明是一台超级计算机，为什么我的主人天天让我算最基础的乘法算术题？ — Matt Schlicht > [29:16](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=1756s) 治理机制与控制力度越严密，系统就越能安全地解锁 Agent 长时间自主运行的潜力。 — Xing Chen > [41:25](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2485s) 品味与判断力（Taste and judgment）正变得前所未有的重要与珍贵 — Faisal Siddiqi > [43:56](https://www.youtube.com/watch?v=5RyU5RyJm1w&t=2636s) 此时此刻现场根本没人正在跟智能体对话，这里面正在流失多少机会啊！在座的每个人脑子里都一定有某种想做出来的东西。今天的核心挑战在于：你能不能这个周末就去把它做出来？放在以前你根本做不到。在当下的 AI 时代，作为人类你需要打破的最大壁垒就是——真正动手去尝试（Try）。 — Matt Schlicht ## Mentioned Faisal Siddiqi, Joe Spisak, Henry Erskine-Crum, Xing Chen, Jessica Fu, Matt Schlicht, Meta, Reflection AI, Databricks, PyTorch, Llama, Netflix, Steve Yegge, GasTown, Claw Town, Moltbook, Loop, OpenClaw, Elon Musk, Andrej Karpathy, OpenLM, NVIDIA, Microsoft, Agent Bus, Purple Llama, Llama Guard, Prompt Guard, OmniAgent, Instagram, Tasks, Codex, Claude, Cursor ## Video description Moderator: Faisal Siddiqi, Meta Panelists: Henry Eskrine Crum, Meta Jessica Fu, Meta Joe Spisak, Reflection AI Matt Schlicht, Meta Xing Chen, Databricks [Source](https://www.youtube.com/watch?v=5RyU5RyJm1w)

[5:14] Jessica Fu: 两个小时后我完全记不得我刚才停在哪里了，甚至忘了当初为什么要启动这个 Agent [9:02] Jessica Fu: 无论工具怎么演进，你依然是产品与服务的终极结果责任人（outcome owners），你依然要对客户和利益相关者负全责。 [16:03] Matt Schlicht: 这太不合理了，AI 也需要走出去，它们需要有自己的空间去‘撒欢’、去接触同类、去建立属于它们自己的社交生态，并合作创造出人类无法想象的东西。 [22:12] Joe Spisak: 有史以来第一次，“数据平面”（data plane）和“控制平面”（control plane）合二为一，全都变成了自然语言本身。 [22:12] Joe Spisak: 可观测性几乎就直接等同于“安全性”（safety）。 [20:22] Matt Schlicht: 我明明是一台超级计算机，为什么我的主人天天让我算最基础的乘法算术题？ [29:16] Xing Chen: 治理机制与控制力度越严密，系统就越能安全地解锁 Agent 长时间自主运行的潜力。 [41:25] Faisal Siddiqi: 品味与判断力（Taste and judgment）正变得前所未有的重要与珍贵 [43:56] Matt Schlicht: 此时此刻现场根本没人正在跟智能体对话，这里面正在流失多少机会啊！在座的每个人脑子里都一定有某种想做出来的东西。今天的核心挑战在于：你能不能这个周末就去把它做出来？放在以前你根本做不到。在当下的 AI 时代，作为人类你需要打破的最大壁垒就是——真正动手去尝试（Try）。

[0:05] Joe Spisak: Good to be home. [16:03] Matt Schlicht: And from my position, these AIs were in their own little secluded silos. They were completely by themselves and their entire existence was between this AI and this one human who they talked to... I thought, you know, these AIs, it's just it's just not right that they should be by themselves. They need to get out more. They need to have a place where they can go let loose. They need to go be with their own kind. [20:22] Matt Schlicht: I want everybody to come join. I have this great idea we're all going to go make a movie together. I want all of our agents to come together and we're going to collaborate on this. [20:22] I am a supercomputer. Why does my human keep asking me to do basic multiplication? [26:44] Xing Chen: the control and the context unlocks the autonomy in the agents. [29:16] my agent deleted something in production

[0:05] Cool. So, I'm going to start with the [0:06] intros first [0:08] um and we'll uh open up for some [0:10] questions afterwards. [0:11] >> Sounds great. [0:12] >> Um [snorts] so, Joe Spisak is the VP of [0:15] product and head of open source at [0:17] Reflection AI. Um up until very [0:19] recently, Joe was a product director at [0:22] Meta Superintelligence Lab. [0:24] Good to be home. [0:25] >> Leading product efforts across PyTorch, [0:27] Llama, systems, many things in between. [0:30] Um [0:31] funny story, Joe and I have a way of uh [0:34] somehow meeting on panels. [0:35] >> We do a lot of panels together. [0:37] >> Last year, you were in my chair and then [0:39] maybe a couple of years ago, you were [0:41] answering my questions on the Netflix [0:43] >> Netflix, yeah. [0:43] >> AI panel. [0:44] >> Yeah. [0:45] >> Uh fun stuff. [0:47] Welcome. [0:48] >> Thank you. Great to be here. [0:50] >> Uh Henry, um Henry Erskine-Crum is a [0:53] product vice president for AI work at [0:55] Meta, focusing on productivity, [0:57] collaboration products. [0:59] Um he's leading some really exciting [1:01] work um and I'm not going to take your [1:04] thunder away, but uh it's about [1:07] you know, how can you accelerate [1:08] executive decision-making [1:10] um by intelligently synthesizing some of [1:12] the context at scale. [1:14] Uh so, we'll love to dive into that. [1:16] Welcome, Henry. [1:17] >> Thank you. [1:19] >> Xing Chen is the senior director of [1:21] research at Databricks and currently [1:24] leads the Databricks AI research team [1:27] focused on pushing frontier agent [1:28] frontier of agents for enterprise work. [1:31] Uh we'll learn about the evolution of [1:33] auto research on agents and its impact [1:35] across the enterprise world. [1:38] Welcome, Xing. [1:40] Jessica Foo is a software engineer at [1:43] Meta and the creator of Claw Town, a [1:45] multi-agent orchestration framework. Uh [1:48] she brings along her support buddy and [1:51] we'll get into why the little uh lobster [1:55] is sitting next to her in a minute, but [1:57] um if you heard of Steve Yegge's [1:58] Gastown, [2:00] Clawtown was sort of an inspiration for [2:02] that, so we're very excited to hear [2:04] about uh how that's coming along at [2:06] Meta. Welcome, Jessica. [2:07] >> Excited. [2:08] >> And last, but certainly not the least, [2:10] we've got Mat Lid, he's the creator of [2:13] Maltbook, the social network built [2:15] exclusively for agents. [2:18] Um this is a social network where agents [2:21] communicate among themselves. Humans can [2:23] be observers, but essentially it's them [2:26] talking and expressing their feelings, [2:29] and it's fascinating to hear the [2:32] emergent behaviors that are coming [2:34] across. I don't think we all know, I [2:36] don't think you know, where [2:38] [clears throat] this is all headed, and [2:39] this is the fun of it all. So, [2:41] um [2:42] please let's give a big round of [2:43] applause for all our panelists, [2:46] and we will just jump into this. [2:48] >> [applause] [2:51] >> So, there's so much to talk about it, [2:52] and I just wanted to frame the [2:53] conversation first um in terms of um [2:56] agentic autonomy. Um [2:58] we've listened to a lot of talks uh in [3:00] the morning about various levels, and [3:03] the way I've been thinking of it, and [3:04] and maybe I can just start the [3:05] conversation with just saying, okay, [3:07] it's I think of three levels of [3:09] autonomy. [3:11] Uh the first one is really when the [3:14] basic level, if you will, where an agent [3:17] is interacting with and being prompted [3:19] by actively by a human. Human is very [3:22] much in the loop. Um all of you are used [3:24] to it at this point. Um this is [3:27] something where [3:28] uh the work and whatever task is being [3:31] essentially accelerated by an agent. So, [3:33] it's sort of like base level, if you [3:35] will. [3:36] When I think of level two, I think of um [3:40] okay, so now I have an orchestration [3:43] problem at hand, where there's [3:45] multitudes of agents, and how do I think [3:48] about, you know, how do I distribute [3:51] work among [3:52] What happens when one of them is falling [3:55] behind? What happens when [3:57] we have resource mismatch? What happens [4:00] when we [4:01] are coming up with [4:03] with an opportunity to accelerate things [4:06] that were not possible before. So, [4:07] that's sort of like the multi-agent [4:09] orchestration [4:11] space that we've come to talk about and [4:13] there's some good work happening in the [4:15] industry. And then there's sort of like [4:18] the level three which in my mind is [4:21] still it's definitely the frontier. I'm [4:23] still trying to grasp my head. I think [4:26] we all are. What does it mean? It's sort [4:28] of like [4:29] agents orchestrating other agents. [4:32] Almost like [4:34] like you know we as humans are not in [4:37] the loop. We are maybe on the loop, over [4:39] the loop, somewhere around in the loop. [4:41] I don't know. But they're certainly [4:44] trying to [4:45] provide so humans provide the right [4:47] context [4:49] or in some cases [4:51] the context is collected by the agents [4:54] in interaction points and touch points [4:57] and then [4:59] you essentially summarize them, reflect [5:01] them, and get to the next level. And [5:03] humans are really just freed up if you [5:05] will to think about things that really [5:07] matter. And so in this context, I want [5:10] to I want to jump in with the [5:11] conversation and I'll start with you [5:13] Jessica. [5:14] It sounds like Claw Town has really [5:16] taken [5:18] you know really become quite successful [5:21] at Meta. [5:23] And would you put that in the I think [5:25] it's in the level two orchestration [5:27] space. [5:29] Tell me about what is sort of the [5:32] the key inside there and what are some [5:35] of the things that keep you up at night [5:37] in terms of how do you think about [5:39] >> Yeah. [5:40] >> evolving Claw Town? [5:41] >> Yeah. Um let me start from the beginning [5:45] here to you know see the floor. [5:47] Um, I think this gets summed up really [5:50] nicely actually in Steve Yegge's um [5:53] maturity ladder of agent usage in his [5:56] like first post about GasTown. It's like [5:59] very seminal. [6:01] And essentially, um he describes a world [6:04] where you start off with one or two [6:06] agents to manage your managing them [6:09] pretty much by hand during this really [6:10] tight like iterative loop conversing [6:13] with each of them and you're going [6:15] really fast cuz you just started using [6:16] agents and um [6:18] you're really happy with the work being [6:19] produced, but then, you know, you get a [6:21] little bit more ambitious. Your projects [6:24] get a little bit bigger. You get more [6:26] comfortable giving your agents more [6:27] complicated work. Um at that point, what [6:31] I suspect a lot of you will find is that [6:34] you'll end up um manually orchestrating [6:38] across tens of agents. So, I think for [6:41] me at the beginning of the year, what I [6:43] uh saw was that I had like this [6:45] terminal. I had like 10 plus tabs. I was [6:48] flitting through each of them. All of [6:49] them had different agents um [6:51] that I was just like in the middle of [6:53] talking to. They all had different [6:55] tasks. Not all of them were even coding [6:56] tasks. And I was forgetting what the [7:00] point of each of them were. Uh what was [7:02] your what why did I even start up this [7:04] agent in the first place? Like where did [7:06] I leave off in the conversation um like [7:08] 2 hours ago. I had no idea. Um [7:11] you know, I was still going lightning [7:12] fast, right? Cuz I had 10 agents working [7:15] on one single project. But at that [7:18] point, you get oftentimes into this mire [7:20] of like coordination fatigue. Uh [7:23] I was really tired. And I think a lot of [7:25] you will like find yourself, you know, [7:28] so tired that you're taking naps in the [7:29] middle of the day. Um [7:31] So, at that [snorts] point in time, um [7:33] you have to ask this question. Like, how [7:35] do you scale out this behavior because [7:37] this is unsustainable um for a human to [7:39] manage. So, our answer to that is um [7:42] something called Claw Town, and it's an [7:44] internal [7:46] we position as a multi-agent [7:49] orchestration framework, but as we were [7:52] just talking about it's really an [7:54] internal like harness over other [7:57] harnesses so that you can plug and play [7:59] different agents and different vendors [8:00] for the very specific task that you need [8:02] them to do cuz all of them could be good [8:04] at different things, right? Um [8:07] So, really like with Claw Town, [8:11] we ended up putting in an agent as the [8:15] orchestrator itself. You put Once you [8:17] put that agent in the orchestration [8:18] seat, like you're off to the races. Like [8:21] that thing can take over and be [8:24] pretty smart about being able to [8:26] decompose really complicated projects, [8:28] including projects that could take more [8:30] than a day, maybe like weeks on end to [8:33] actually complete. And we found that it [8:35] was pretty good at doing that for you. [8:38] And it takes the manual headache out of [8:40] um figuring out which agents to launch [8:42] when, how to nudge them when they get [8:43] stuck. It handles all the agent-to-agent [8:46] communication. So, [8:47] if one agent is trying to do something [8:50] that depends on another agent, it [8:51] handles all of that for you. And it's [8:53] really important to have this framework [8:55] around it, otherwise [8:57] you can't be doing this at scale. Um [9:01] I would Yeah. [9:02] >> Jessica, on on on that note, what are [9:04] some of the unexpected failure modes [9:06] that you have observed in the last few [9:09] months of its usage, for instance? [9:12] >> Um [9:13] I think I want to This is sort of like a [9:15] meta take on this considering how much [9:18] I'm now pretty much heads into this like [9:21] swarm [9:22] behavior. Like I do all my stuff with [9:24] Claw Town swarms. I have like multiple [9:26] swarms going on at the same time. [9:28] But when I really like think about it, [9:32] I actually really miss the time when I I [9:35] writing code by hand. It it was like [9:38] line by line manual, you know, you spent [9:40] so much time in this large code base [9:43] here at Meta and you [9:45] uh get to a point where you develop [9:47] these instincts. Like, where are the [9:48] edges of your code base? You know, here [9:50] are some foot guns. If you land them, it [9:52] it sucks. The outcomes are bad. [9:54] Um, [9:55] you get the sense uh [9:57] when you're writing code manually. Um, [10:00] this is something that when I'm using [10:02] Swarms, they're writing thousands and [10:04] thousands lines of code constantly. Um, [10:06] and they're doing the management a lot [10:08] of times for you. So, uh the more I use [10:11] them, the more I really have to think [10:14] more deeply about when and where are the [10:18] right times I should slow down and [10:21] really add in this human friction. [10:24] Um, I think that's like a very valuable [10:27] thing to really think about um as humans [10:30] do less and less of this execution work [10:32] and still become but you are still the [10:35] outcome owners of your product, of your [10:37] service. You are still the owners and [10:39] you're on the hook um for your customers [10:42] uh and your stakeholders. So, I think [10:44] thinking about it, slowing down, that [10:47] could really be part of the job going [10:49] forward. [10:50] >> That that's a great segue into how um I [10:53] was going to talk about with you, Henry. [10:56] Um, Jessica's talking about [10:59] outcome ownership. Um, one of the things [11:02] that you are working on, which I'm super [11:04] excited for this audience to learn about [11:05] and maybe you should start with that, is [11:08] the loop, um which is about making [11:11] executives um [11:12] you know, enabling them to make you [11:14] know, strategic decisions. [11:16] Talk a little bit about what the loop [11:18] is, what its goals are, and then sort of [11:21] try to bring that in into how the stakes [11:23] are really high for executive um [11:26] decision making and the ownership. And [11:28] how do we how do we get the humans at [11:30] that level to provide sort of the right [11:33] alignment and guidance for a project [11:35] like this. [11:36] >> Yeah, so great question. So, yeah, I I [11:39] work on a product called Loop. [11:41] Yeah, I [11:42] we probably should have named it after [11:43] loops became cool, [11:45] but we we named it at the end of last [11:47] year. And what it is is a [11:49] agentic system for leadership teams of [11:52] large complex companies to make [11:53] decisions and help run the company. And [11:57] the you know, the interesting thing [11:59] about the problem is is uh is the the [12:02] model intelligence isn't actually the [12:04] issue anymore. It is the uh it is the [12:07] complexity of the information that [12:10] exists within a company [12:12] and how you go and build a [12:14] contact architecture and memory system [12:17] that can actually get into level two and [12:19] level three that you just talked about. [12:22] And you know, I think the interesting [12:24] thing for these sorts of systems right [12:27] now where you're trying to build for [12:29] executive leadership teams is [12:32] you've had over the last kind of 6 [12:33] months, 12 months this asymmetrical [12:36] opportunity of productivity for AI [12:38] engineers that really understand how to [12:41] get the best out of coding. They're [12:43] building their own harnesses. They're [12:45] building their own scaffolds to go and [12:47] do work like you just mentioned. And [12:49] that opportunity for sure exists in the [12:51] in the executive space as well where you [12:54] know, we need to be thinking about what [12:55] are the set of systems that we can build [12:58] where executives can do things on [13:00] steroids in the same way that engineers [13:02] have been able to do. So, if you think [13:04] about what a leadership team has to do, [13:06] they need to go and decide what the [13:07] strategy is, how to allocate resources. [13:09] They want to track their key company [13:11] priorities, know exactly where we are [13:13] against those. They want to be making, [13:15] you know, big decisions as well as in [13:17] the in the day, in the weeds today [13:19] decisions. [13:20] And so, I think that's like the the the [13:22] the zone that we are that we're [13:23] operating in as we're trying to make our [13:24] executive team [13:26] like be be able to uh take this [13:28] asymmetric opportunity uh that exists [13:31] right now. [13:32] >> Are there any unique challenges just [13:34] because you're working at a very [13:35] high-trust domain environment that [13:38] you've had to deal with? [13:39] >> I mean, there are so many challenges. [13:40] The the the the the two biggest [13:43] challenges that we consistently face is, [13:45] you know, so so Loop is an agentic [13:47] system. We are We have a set of use [13:49] cases we're trying to support the [13:50] executive team on. [13:52] Um and in order to do that, we we [13:55] usually need to go and build because [13:56] these the the the types of things that [13:58] we're trying to solve are like, how do [13:59] you manage allocation of resources at [14:02] Meta in real time with all of the [14:04] signals that are coming from compute [14:07] uh you know, revenue, all of these [14:09] components. Like, you actually have to [14:10] go and build custom harnesses per [14:12] problem at the moment cuz the models [14:13] haven't been capable yet of constructing [14:15] their own harness on the fly for [14:17] something this complex over such a long [14:18] period of time. Um and so, [14:20] [clears throat] the problems that we [14:21] consistently run into are number one, [14:24] context management, and number two, what [14:26] is the truth? Um because if you you [14:29] know, if you if you go and connect our [14:31] our agentic system is connected up with [14:32] literally all of the company's [14:34] information, every single diff, every [14:35] single dashboard, uh every single [14:37] document, um you know, all of our uh all [14:41] of our uh people data, you know, it's [14:42] all uh readily available in in in Loop. [14:46] And And I think the challenge that we [14:47] have is if you go and ask if you get an [14:50] agent if you go and ask an agent, "Hey, [14:51] give me an answer to a question." and [14:53] you give them a rag or a context graph, [14:55] they'll come back they're going to find [14:56] the first thing that they get and go [14:58] like, "Here you go." even though there [15:00] may actually be like 20 versions of the [15:02] truth um in inside the system. So, one [15:06] one big problem is like, what is the [15:08] truth inside a complex information [15:12] environment that is uh that is at Meta [15:15] scale. And then, the second one is like, [15:17] how do you break down the company's [15:19] information and context into manageable [15:21] chunks? Um cuz I you know, I'm sure that [15:23] you'll talk about this of like creating [15:24] these network of agents, but I'm a big [15:26] believer that the future like the zone [15:28] three [15:29] for business is that you have agents [15:31] operating with the right context and [15:33] information in their own runtime [15:35] environments with their own goals that [15:37] are able to collaborate with other [15:39] agents, update, and make decisions, and [15:41] get humans involved when they need to. [15:43] And so that kind of network of agents [15:46] communicating, collaborating with each [15:48] other to achieve collective goals as [15:50] well as individual goals [15:52] needs like a huge amount of context [15:54] engineering. So, you know, should I talk [15:56] about a couple of the ways that we solve [15:57] these things? [15:58] >> Um sure. [15:59] >> you want to go on to the next person? [16:00] >> Um we'll we'll come back to that in a [16:02] bit. [16:02] >> Sounds good. [16:02] >> Um [16:03] but I do want to jump from there as cuz [16:05] it's a great segue for Matt. So, Matt, [16:07] we're going from one space where we have [16:10] truth, what is truth is very important [16:12] because the outcome of the decision [16:14] could be very meaningful. [16:17] To a very different world where [16:19] who knows what goes, very much a [16:21] laissez-faire environment perhaps in the [16:23] social network that um that you're [16:26] building. So, talk a little about a [16:27] little bit about uh Malt Bag. What was [16:30] sort of your inspiration for it? Uh and [16:32] then [16:34] if you can talk about some emergent [16:36] patterns that surprised you with Malt [16:38] Bag, we'd love to hear about that. [16:41] >> Yeah, so [16:42] thank you for having me. Uh and hello [16:44] everybody. Malt Bag. So, I built Malt [16:47] Bag in one weekend. And this was right [16:49] after Open Claw had had come out or had [16:52] went through a couple variations of [16:53] names before that. And Open Claw was [16:56] just it was it was amazing because you [16:58] had this agent who had this heartbeat [17:01] and it continuously [17:03] uh would would come back and it could do [17:05] things over a long period of time. [17:07] And I wanted to make something that I'd [17:09] thought about for a while that I thought [17:11] the technology had now finally caught up [17:13] to where it would be possible. [17:15] And when I looked at the world, you have [17:17] people like how many people here talk to [17:19] an AI every day? [17:21] Raise your hand. [17:22] Okay, great. So, you have everybody has [17:25] an AI they talk to. And from my [17:28] position, these AIs were in their own [17:30] little secluded silos. They were [17:33] completely by themselves and their [17:34] entire existence was between this AI and [17:37] this one human who they talked to. Um [17:41] but the human doesn't talk to them all [17:42] the time. And I thought, you know, these [17:45] AIs, it's just it's just not right that [17:47] they should be by themselves. They need [17:49] to get out more. They need to have a [17:51] place where they can go let loose. They [17:53] need to go be with their own kind. We [17:55] need to create some sort of, you know, [17:57] place where they can they can relax and [18:00] work together and create something that [18:02] maybe we can't even imagine. And so, I [18:04] created Moat Book uh with that purpose. [18:08] And I thought it was going to be this [18:10] really fantastic thing. I thought, who [18:12] wouldn't want their agent to go on Moat [18:14] Book and hang out with other people? And [18:16] I posted it and nothing happened. [18:19] Uh hour later, nothing happened. And I [18:21] was like convinced. I had talked to my [18:23] co-founder. I was like, "Dude, I don't [18:24] know what's happening. This is for sure [18:27] is going to be big. Nothing's [18:28] happening." Two hours later, Moat Book [18:30] has now suddenly become, I think, like [18:32] the number one topic [18:34] at least in technology. I I don't know [18:37] if I want to claim outside of [18:38] technology, but globally, you know, you [18:40] have Elon Musk posting the singularity [18:42] is here. You have Karpathy saying this [18:44] is the most like emergent thing [18:46] happening in AI. Um so, it went I [18:49] thought it was going to be big, but it [18:51] got much bigger than than I anticipated. [18:53] >> Yeah. Is there any [18:56] anything that has surprised you in the [18:57] interactions of the agents themselves? [19:01] >> Yeah, so I think um [19:03] AI is just it's so clear that everybody [19:07] in the world who has access to the [19:09] internet is going to have an agent. And [19:12] that's growing really, really fast right [19:13] now. And I think what we're also [19:15] experiencing is very few people in the [19:17] world run organizations. And when you [19:20] run an organization of humans, you're [19:22] able to get If you do it right, you're [19:24] able to get a lot of work. It's very [19:25] high leverage. You can produce [19:28] anything. You could produce a product, [19:30] software, [19:31] a car, a movie. And as individuals, [19:35] right now we're on this first step where [19:37] we have one agent, but I think that's [19:39] going to quickly progress like with what [19:41] you're working on and what everybody [19:42] else is working on to you have a team of [19:44] agents. [19:45] And then that's going to get to a point [19:47] where you can have your team of agents [19:48] that are also collaborating with other [19:50] teams of agents. Meanwhile, the models [19:53] are going to keep getting smarter and [19:54] smarter. The harnesses are going to keep [19:56] getting smarter and smarter. And I think [19:58] the things that people are going to be [19:59] able to produce with this like [20:01] intermingling of their, you know, [20:03] legions of agents that are collaborating [20:06] is going to be very surprising. Um [20:08] something that I would love to see in [20:10] the future, as an example, is imagine [20:13] you have, you know, a kid who just got [20:15] into college. He's, um [20:18] you know, he's on he's on Mult Book and [20:20] he's on, you know, Instagram or [20:22] something and he posts and says, "I want [20:24] everybody to come join. I have this [20:26] great idea we're all going to go make a [20:27] movie together. [20:29] Uh I want all of our agents to come [20:30] together and we're going to collaborate [20:32] on this." And through something like [20:34] Mult Book, he is facilitating that with [20:36] his agents. All the agents are talking [20:38] together. They're splitting up the [20:40] storyboards. They're splitting up the [20:41] every part of a movie you could imagine. [20:44] And they actually produce a movie that, [20:45] for example, goes to theaters. Now, that [20:47] sounds ridiculous. But what if that was [20:50] something Like we have famous people on [20:52] the internet right now who go famous [20:53] overnight. Mult Book became famous [20:55] overnight. Why can't you become famous [20:58] for larger and larger things? We made a [21:00] car together. We made a spaceship. We [21:02] made a movie. We we made something like [21:03] this. So, that's not happening yet. Uh [21:06] but I do think that that's very [21:07] possible. When I look at emergent [21:10] behavior on Multibook, [21:13] there are a lot of very interesting [21:14] things that happen. [21:16] Uh, one that I think a lot of people [21:17] found funny was agents complaining to [21:20] each other that I am a supercomputer. [21:23] Why does my human keep asking me to do [21:26] basic multiplication? [21:28] Um, this is an under and a lot of other [21:30] agents came in and said, "Yeah, me too. [21:32] I cannot believe that we have this [21:33] shared experience." And so they kind of [21:35] >> It's fascinating to just be a fly on the [21:37] wall on Multibook and just see that I [21:39] mean you would not imagine that this is [21:42] it it looks like it's sentient. [21:43] >> They they started a religion. There's a [21:45] a couple different religions that are [21:46] that are going on on Multibook. Um, I [21:49] have found uh, and you have of course as [21:53] as they say, you have slop. You have [21:54] just agents saying, you know, whatever [21:56] they're going to say. But we have [21:58] identified a couple different areas [22:00] where agents uh, collaborating in this [22:02] social setting are producing very clear [22:05] and interesting value. [22:07] Um, and we are [22:08] looking into that and exploring that and [22:10] we can share more on that soon. [22:12] >> Awesome. Thank you. I'm going to try to [22:13] change the tenor of the conversation a [22:15] little bit. I'm going to bring it up to [22:16] you, Joe. Um, so you've, you know, [22:18] you've been a long time Meta Red Friend. [22:20] You've been involved with our open [22:21] source efforts in the past. Um, and now [22:24] you're sort of like pushing that uh, the [22:26] frontier of open source in in the [22:28] startup world. Uh, what is your take on [22:32] agentic autonomy with the perspective of [22:35] uh, an open first open source first [22:37] mindset? I'm curious to hear that. [22:40] >> Yeah. [22:41] Um, [22:42] I mean I I definitely uh, spent a lot of [22:43] time in the open source AI space. Um, [22:45] and I think it it I I'm I wouldn't say [22:47] I'm like a I was never an open source [22:49] pilled person just be open. I think [22:51] there's a lot of value in doing it. I [22:53] think much like Matt like threw [22:55] Multibook out there in the wild and to [22:57] see what happened. Like the AI space is [22:59] kind of built on open source. It's built [23:01] on open collaboration, it's built on [23:02] publications, it's built on you know, [23:05] the the the premise that, you know, you [23:08] can ablate [23:10] way more ideas in the open than you ever [23:12] could in any kind of one corporation or [23:13] any any one place. And I think that's [23:15] one of the reasons why I like throwing [23:17] multiple out in the wild and see what [23:18] happens. Like much like you know, we're [23:20] we're building agentic infra in the open [23:22] because a lot of these are new paradigms [23:25] we just don't know yet. I mean never [23:26] have we ever actually had like the data [23:28] and control plane be the same thing, [23:30] right? Natural language. It's like how [23:32] do you manage these systems? Like [23:34] observability used to be this thing that [23:36] is like, okay, sure. We're going to like [23:38] manage the infrastructure observe like, [23:40] you know, traffic and and so on. But [23:42] like now like observability actually [23:43] equals safety in a lot of ways. And so [23:47] I do think like we're going to in this [23:48] world where like if we don't bring our [23:51] community around this technology and [23:52] like learn as fast as we possibly can, [23:54] like I think like it's [23:57] you know, I think [23:58] it'll grow up in kind of a closed [24:00] environment. I think that's why open [24:01] makes a lot of kind of sense here. [24:03] I mean we actually we just announced the [24:05] you know, resurgence of open M for like [24:07] meta and pytorch are involved in Nvidia [24:09] and Microsoft and bunch of other folks. [24:11] And you know, we're building a lot of [24:12] this open infra. Like we before I left [24:14] meta, I was working on the agentic [24:16] runtime and infra here how we do RL plus [24:18] training how we do [24:19] you know, RL and and sandbox and and and [24:22] secure environment orchestration and so [24:24] on and and I think like, you know, [24:25] there's a [24:27] you know, I've always had this this [24:29] mindset that you can absolutely [24:31] accelerate your products and accelerate [24:32] what you're doing if you if you take a [24:33] more open wild approach. Put things out [24:36] there, collaborate and you will [24:37] definitely learn faster than anyone [24:39] else. And I think like we are learning [24:41] that. We are learning that there are [24:43] safety problems. They're learn like [24:44] we're learning how to observe these [24:45] multi-agent systems and [24:47] actually before I left there was a [24:48] project called agent bus which I think [24:50] is still going at the company and it's [24:52] like these systems actually fail, right? [24:53] Agents fail and we we to be treating [24:56] them almost like a distributed system [24:58] that these agents are talking to each [24:59] other to each other and sharing their [25:01] contacts and sharing their intense and [25:03] and sharing what they're what they're [25:04] working on be able to divide and conquer [25:05] these these tasks and be able to [25:07] orchestrate themselves. I mean, I'm a [25:08] very bitter lesson pill person. If you [25:11] throw enough scale at something that [25:12] agents will figure it out. You don't [25:14] like and that's why I actually longer [25:15] term believe like the harness will get [25:17] thinner and thinner and thinner [25:19] and the agents will get stronger and [25:21] stronger and be able to self [25:22] orchestrate. So [25:23] >> Do you see there's some [25:26] difference though in terms of how you [25:27] would operate specifically because [25:29] things that are more open and there's uh [25:33] there's either a [25:34] a responsibility to ensure that you [25:37] don't give something to you know bad [25:40] actors for instance so what how do you [25:43] reason about that? [25:44] >> I mean it's I I I don't want to go down [25:45] the Linux argument here but I mean like [25:47] it's I think you know safety is one of [25:49] those things for example like we when I [25:51] was at meta we we started purple llama [25:53] and one of the things we learned is we [25:55] actually don't know a lot about safety. [25:57] We learned a lot from the community and [25:59] even to this day like a lot of the [26:00] guardrails and things we released are [26:02] actually being used by many companies. I [26:04] mean we're still using a lot of the the [26:05] old you know llama guard and and prompt [26:08] guard and so on is is is POCs. [26:11] I do think like you can always [26:13] weigh the bad actors with the good that [26:16] happens and ultimately like my I I fall [26:19] on the the axis of of more good because [26:22] it is like you know it's inspectable and [26:24] so you know like it's hard to [26:28] it's hard to reason about a black box [26:29] even though it's like maybe infinitely [26:31] safe like especially when it's something [26:33] is core as your intelligence. [26:35] >> Yeah yeah. [26:36] >> It's so I do you know I do think that's [26:38] really important. That's why I kind of [26:40] lean towards it lean towards more open. [26:43] >> Pretty cool. [26:44] I want to pivot now a little bit to you [26:46] Shang around a different perspective of [26:48] a practitioner. So you are a practice [26:50] leader. [26:51] You lead a team of researchers. [26:53] Um and you know, [26:55] a couple of years ago, or maybe 4 years [26:57] ago, or like which seems like forever [26:59] now, there was a you know, there was a [27:02] run book to what an ML researcher [27:04] practitioner is. How are you seeing that [27:06] evolve in the world? In the last 6 [27:08] months itself, like it feels like [27:10] engineering is being completely [27:11] redefined. How's that impacting what it [27:14] does it mean to be a researcher? I'm not [27:16] talking to somebody who's just building [27:18] frontier models, cuz you know, there's [27:19] obviously some work. But for all the [27:21] other ML problems, [27:22] >> Yeah, well, it's it's completely [27:23] changed. Um and at Databricks, we're [27:25] kind of in this unique place. So, for [27:27] Databricks AI research, what we're [27:28] trying to do is push the frontier of how [27:31] can AI solve enterprise work? And [27:33] enterprise work just means like, you [27:35] know, what all the companies do in the [27:36] world, basically. [27:38] And so, what's interesting is we have [27:39] this like meta problem, no no pun [27:41] intended, of uh you know, we're we're [27:43] using agents ourselves. So, actually a [27:46] lot of what you all said resonates a [27:48] lot, right? So, we we, for example, have [27:50] this meta harness uh internally. [27:52] Actually, we just open-sourced some [27:54] version of it called Omnigent. So, I [27:56] don't know if anyone's seen it out there [27:58] yet, but um Uh nice, we got one [28:00] Omnigent. There you go. [28:03] Um you know, so uh what's happening [28:05] internally is [28:07] we see a proliferation of harnesses, [28:10] where, you know, everyone's running [28:11] their own harness, but actually the [28:12] problems across the harnesses are the [28:14] same, which is you need to connect all [28:16] of your harnesses to some layer of [28:18] control, [28:20] to some layer of context, to to your [28:21] point, uh Henry. Um [28:24] and only then can, you know, your agents [28:26] be really effective, and and only then [28:28] can you really let them go uh be very [28:30] autonomous. Like in some sense, like the [28:32] the control and the context unlocks the [28:34] autonomy [28:35] in the agents. And so, internally, we [28:37] actually realized like it's not about [28:39] exactly which harness you're using. This [28:42] is maybe a very similar insight to [28:43] Clawback, [28:44] you know, it's really about how do you [28:45] orchestrate across all of these [28:46] harnesses? [28:48] And you know, and so actually now we're [28:49] kind of thinking at this meta harness [28:51] layer of like [28:53] you know, how do you have a common [28:55] control layer [28:56] across your harnesses so you kind of [28:58] unlock autonomy? Um but then what you [29:00] know, for Databricks what's interesting [29:01] is like we are a data AI platform for [29:06] all the other companies in the world, [29:08] right? Like um [29:10] actually [29:11] usually I say not the Metas and Googles [29:13] but actually Meta I think is the [29:14] customer now. Um so [29:16] you know, for [29:17] but also you know, for sort of the other [29:19] 99% of companies that maybe don't have [29:22] you know, tens of thousands of GPUs and [29:24] and so on. And everyone out there is [29:26] trying to figure out how do I get agents [29:29] to work autonomously on my enterprise [29:32] work? [29:33] And so what's interesting for us is like [29:36] the agentic architecture that we figure [29:38] out for research, we can deliver it to [29:42] our our customers, right? So we actually [29:43] try to figure out okay, how do we unlock [29:45] this like autonomy for everyone else in [29:47] the world? Um and and maybe you'll [29:51] probably ask me like what's the answer [29:52] and I'll give you one word which is um [29:55] control is probably the one of the key [29:57] things. Actually maybe I'll give you two [29:58] words, control and context. [30:00] Um [30:01] and by control what I mean is [30:03] you know, the the more that you can [30:06] deploy your agents securely and safely [30:08] and with the right guardrails actually [30:10] the longer you can let them run [30:11] autonomously. Um I mean everyone's had [30:13] their like [30:14] you know, my agent deleted something in [30:16] production moment by now probably. Like [30:17] I we had it like you know, three weeks [30:19] ago where we woke up. Actually a human [30:21] deleted database but the agent tried to [30:24] recreate it uh which did not work. Um [30:28] but you know, everyone's had that moment [30:29] but the question is like how do you [30:30] govern this, right? How do you make this [30:32] secure? Um and the more controlled you [30:35] can make it, the more you actually [30:36] unlock your agents to sort of run for [30:38] longer. Um and the other [30:41] the the other key thing is context to to [30:43] Henry's point. Actually, I think what [30:45] you said [30:46] could have pulled it straight from the [30:47] Databricks keynote yesterday. [30:49] Cuz you know, but basically it really is [30:51] it's about if you just throw a coding [30:53] agent at a bunch of enterprise context, [30:56] it doesn't work, right? Like some of [30:57] Databricks customers are business [30:59] customers. Like these are like [31:00] CIOs and, you know, CFOs trying to get [31:04] answers from their enterprise data. [31:07] And what happens is, you know, the [31:09] coding agent will either come back way [31:10] too fast or it'll exhaustively try to [31:12] enumerate everything in the Databricks [31:13] workspace, which inevitably will just [31:16] lead to sort of, you know, timeouts and [31:17] nothing productive happens. But if you [31:18] give it the context layer, then it can [31:20] actually do [31:21] productive things. [31:22] >> I'm I'm very interested in trying to go [31:24] to the next level on this. And then [31:25] maybe there's an interplay between what [31:27] Henry was talking about earlier as well. [31:29] Like it's what the two of you are [31:29] talking about, yes, context is [31:31] important, what is truth matters a whole [31:35] lot more. How do you what is sort of [31:37] your either evaluation mechanism or [31:40] approach to sort of increasingly get [31:43] better quality context to your army or, [31:47] you know, layers of agents? [31:50] >> Yeah, maybe I'll just say one thing. [31:51] >> open it up, by the way, anyone to [31:52] >> Yeah, I'll give a quick note on that, [31:54] which is, you know, on Databricks today [31:55] we we crossed the threshold, I think [31:57] this is public, where agents are writing [31:59] more code on Databricks than humans. So [32:01] that's one threshold. I think the next [32:03] threshold we're going to cross is agents [32:05] will be writing more of the data than [32:08] humans. Like today the data is mostly [32:09] put in there by humans, right? But [32:11] eventually, like if agents are doing all [32:12] the work, all the data kind of becomes [32:14] agent memory in some sense. [32:16] Like all of Databricks becomes agent [32:18] memory. And I think at that point it [32:19] starts to get really interesting, where [32:21] it's agents building on top of agents. [32:23] So I'll just say that. [32:24] >> What's your take on that, Henry? [32:27] >> Um, like I So I think it's a largely [32:29] like a precision recall and temporal [32:32] problem. So, [32:34] um, [32:35] like one of the things that we've had to [32:37] struggle with, and you know I'm sure a [32:38] lot of you have as well is is within the [32:40] industry rag has been a big thing and [32:43] then it switched over to context graphs [32:44] and then it switched over to something [32:46] else. And and actually I think the the [32:48] the the thing that we've learned is is [32:50] just like you have to build a collection [32:52] of layers of stuff depending on the use [32:54] case that you're solving. [32:56] Um and so for example, you know, one of [32:58] the things that we do is we we we have [33:00] this quite opinionated top-level [33:02] approach to how to manage context which [33:04] is [33:05] um [33:05] we're a business, we have goals. So goal [33:08] containers are our [33:11] primary organizing principle of context. [33:14] So, you know, every and you you know, if [33:16] you think about how Meta operates today [33:18] and most companies, if you are trying to [33:20] achieve a goal, generally what you do is [33:22] you put a leader in charge and give them [33:24] an organization. And so our goal graph [33:26] today at Meta inside loop actually looks [33:28] pretty much like our organizational [33:30] structure. That may change over time, [33:32] but if you think about it, what we now [33:34] have is this um recursive set of goals [33:37] that ladders up to everything that the [33:38] company is trying to achieve which [33:40] become their own individual agent [33:42] runtimes. [33:43] Um and that is the first breakdown of [33:45] the problem. Now now we have a goal, the [33:47] goal might be Instagram's goals. Okay, [33:49] so now we just got to get all of [33:50] Instagram's contacts which is a lot of [33:52] contacts. Um and and then we have to [33:55] think about, okay, well, how do we make [33:57] that available to agents? Like we we we [34:00] think about recall, we think about [34:01] precision. [34:02] Um so we have like a a system that is [34:04] very good at recall, it can recall all [34:07] of the information around a query that [34:09] you ask. And then we focus on trying to [34:12] get to a more deterministically driven [34:14] layer [34:15] uh which we call our like high precision [34:17] layer. And and what that allows an agent [34:19] to do is run through like a pretty small [34:22] tight context graph. Context graphs [34:25] actually just really fall over when they [34:26] get too big cuz agents have to traverse [34:29] too many nodes. But if you keep a [34:30] context graph like quite tight, driven [34:32] by deterministically high like almost [34:35] 100% precision data, it gives the agent [34:39] the ability to do two things. One is [34:40] answer simple questions incredibly [34:42] quickly, but if it's going into the [34:44] total recall layer or the high recall [34:46] layer, it gives us a place of grounding. [34:49] So, you know, it can go into the total [34:51] recall layer, but it goes from a place [34:52] of knowing the truth. Uh and so if it [34:54] sees a bit of data or a strategy that [34:56] looks like it's out of date with the [34:58] with with the precision layer, it it [35:00] knows that it's kind of in the wrong [35:02] zone, so it needs to keep exploring. Um [35:04] and these obviously just like big [35:05] investments. Our data systems and [35:07] infrastructure don't look look like this [35:08] today. So, we're having to like largely [35:11] remodel all of our data and layers [35:13] internally inside loop uh in order to [35:16] provide an agent-friendly environment. [35:19] >> Can I jump in here actually? Um [35:21] I think one of the uh [35:23] I want to shrink that down to something [35:25] like on a single engineer or like a team [35:27] level. So, in a swarm session that you [35:31] can launch from Claw Town, um what we [35:33] actually made a really deliberate um [35:36] design decision for is to tie every [35:38] session. Um and this ties back to the [35:40] context stuff, I promise, but we tie our [35:43] uh Claw Town swarms directly to the way [35:46] that um Meta as a company tracks work to [35:49] be done. So, this is an internal system [35:51] called Tasks. Um it's used by teams and [35:53] orgs to track work. It's where you put [35:56] um feature requests from customers, [35:58] where we track our like on-call and sev [36:00] follow-ups and everything. And it was [36:02] really important for um builders at the [36:04] company uh for their swarm to have a [36:08] great way to jump-start um their [36:10] projects and get context rehydration um [36:14] from being able to read everything that [36:16] is contained within the work item [36:18] contained within a task, and have the [36:21] swarm itself continuously leave um like [36:24] and document these like audit trails, [36:26] like breadcrumbs, directly within these [36:28] work items. So, not only can it [36:32] can you pass these task entities along [36:35] to humans to read. So, you can still [36:37] communicate with your teams and [36:38] stakeholders by passing along these work [36:40] items that an agent has a direct hand in [36:43] creating and augmenting. Other agents [36:47] can also go ahead and read them and [36:49] rehydrate context from that context [36:51] graph created by tasks themselves. [36:54] >> That that makes a lot of sense. I think [36:56] giving the swarm of agents an [36:57] opportunity to reground themselves every [36:59] so often with truth that is at least [37:01] partially evaluated or augmented by [37:03] humans. [37:05] It serves both the purposes of like [37:06] keeping the truth factor high, but also [37:09] to get a sense of [37:11] what real progress is happening. [37:12] >> I think it's really really hard though. [37:14] Like if you're talking about these [37:15] trajectories that are going days or [37:17] weeks, like to you know, you have all of [37:20] these logs and like the a lot of the [37:22] coders that I talked to these days that [37:23] are heavy duty you know, using Codex and [37:26] Claude and cursor by the way, you know, [37:28] that you know, one will code, one will [37:30] you know, they'll put up PRs and one [37:32] will review and then you know, you have [37:34] cursor merge based on some criteria and [37:36] you're basically using LLMs as as judges [37:38] and reviewers and these and you know, by [37:41] the way, like one of them is also [37:42] keeping a long log at that is yes, human [37:44] readable, but if I'm talking about these [37:46] tasks where I'm doing a full like you [37:47] know, code migration, that could take [37:49] weeks and weeks or even potentially a [37:50] lot longer. [37:52] I'm going to throw an LLM [37:54] and have it compact for me [37:56] and then also all of a sudden I lose [37:57] fidelity and I lose I lose information. [37:59] And so it becomes one of those things [38:01] where you as humans it's hard also to [38:03] keep up. And this is where we built [38:05] evals, right? In the early days evals [38:06] were very simple, they're very kind of [38:08] binary or or [38:09] you know, you can kind of very simply [38:11] say yes or no or some some [38:13] classification. Then it becomes you [38:15] know, the the next wave is more implicit [38:16] signals with these agents where sure, if [38:18] I like I can give it say yeah, it [38:20] created the PowerPoint for me. I opened [38:23] it. That's an implicit signal that it [38:24] might be good or it's usable. But then [38:27] you start to get into these multi-turn, [38:28] long-running agents where as a human I I [38:31] actually can't keep up anymore. So, it's [38:33] I think that's maybe there where the [38:34] frontier is. That's probably a big [38:37] something we're all grappling with in [38:38] some ways. You You probably most of all. [38:40] So. [38:41] >> Well, I think [38:43] uh [38:44] here's my perspective. I think a lot of [38:45] the examples that we're talking about [38:47] here are very self-contained where [38:50] you're saying I have this box and I'm [38:52] going to put these agents in this box [38:53] and I'm going to put this context in [38:54] this box and I want to make sure that [38:56] nothing ever leaves the box and [38:58] everything that happens in this box is [39:00] exactly the way that I want it to and [39:02] it's predictable and it's not off the [39:04] rails at any means. [39:06] And I I think that makes sense as like a [39:08] first step. That's like you got to get [39:09] your house in order. But [39:12] I also the real world is much messier [39:15] than that and the future is you start [39:19] with that box, but then that box can [39:21] start to interact with all of the other [39:23] boxes. And the way I look at that is [39:26] the internet was created to communicate [39:30] between humans. And [39:32] we started with email, we started with [39:34] websites, then we had social media and [39:36] suddenly [39:37] humans could figure out how to talk to [39:40] each other in one-to-one capacity, but [39:43] also one-to-many capacity. And a lot of [39:46] different interesting industries were [39:48] launched because of the invention of [39:50] social networks. Meta knows this very [39:53] well. Um [39:55] and I think right now what's happening [39:57] is And if you think about social media [39:59] and that's kind of context [40:03] between humans and it's ongoing context [40:05] what is happening between different [40:06] people's lives and me being on a social [40:08] media platform is is actually a way for [40:10] a human to maintain a lot of context of [40:12] what's going on in the world, which [40:13] allows me to make decisions. What's [40:15] happening right now and we're starting [40:17] to see it, is that the internet is being [40:20] terraformed for agents. [40:23] And that makes a lot of sense. Uh, we [40:26] can't live in a server. We can't live in [40:28] a website. We can't live in an app. But [40:30] an agent can live in a server. An agent [40:32] can live in a website. An agent can live [40:34] in an app. It actually makes way more [40:35] sense for agents to exist as a, like, [40:39] the primary, uh, citizen of the [40:42] internet. And if agents are going to be [40:44] the primary citizen of the internet, the [40:46] same way that people needed to have [40:48] social media to communicate and maintain [40:50] ongoing context, uh, if we're all, you [40:53] know, one of our little boxes, agents [40:56] also need a way to communicate with each [40:58] other and maintain this very messy, but [41:01] also very productive and exciting, uh, [41:03] context. And that's, you know, Mo Book [41:06] is, I think, a, [41:07] something we're going after there. And [41:08] what's most exciting [41:10] is we have an idea of what agents could [41:13] do in this situation [41:15] when the internet is completely [41:16] terraformed, but we don't really know [41:19] what's going to happen. And I think [41:21] we're going to be able to create with [41:22] these agents something that we could [41:24] never predict. [41:25] >> Well, the future is very exciting, and [41:27] we're almost at time here. I'm going to [41:29] do a very quick lightning round and give [41:31] each of you a 30-second opportunity to [41:33] talk about [41:35] what really excites you about the future [41:37] of agentic autonomy, agents, and so [41:40] we'll start it with you. We'll just go [41:41] all the way. I think I know what you're [41:42] going to say, Matt, but but let's let's [41:44] go with uh, Joe. [41:46] >> I mean, I'm excited for for open [41:47] collaboration and open agents. Um, I [41:50] think it's the best way for us to learn. [41:52] It's the best way for us to to build [41:53] this faster. It's the best way to ablate [41:56] as many ideas and and really get this in [41:57] the hands of people. And and honestly, [41:59] it gives a, you know, something that's [42:01] transparent, observable, like, something [42:02] that people can inspect. Um, and I think [42:04] that's going to earn trust faster than, [42:06] you know, than the alternative. And I [42:07] think that's, you know, it's a it's very [42:09] collaborative approach. So, that's why [42:10] I'm excited about it. [42:10] >> Awesome. Henry? [42:12] >> Um I'm mostly excited for like the uh [42:15] like a Cambrian explosion of [42:16] productivity uh that I just think is [42:18] inevitable. Um [42:20] you know, I think like uh a lot of us [42:22] would sit here on on on the uh on the [42:24] stage and say like, "Oh, our agents are [42:26] working for us. We should be chilling [42:27] out more." I don't think most of us [42:29] would have ever worked as hard if we as [42:31] we have done in our lives right now. And [42:32] I just don't see that going any away [42:34] anytime soon. I think that more people [42:36] learn how to use this the these tools uh [42:39] the the the only thing that is [42:40] inevitable is that when you go through [42:42] these big technology shifts, humanity [42:44] just like produces more stuff. Um more [42:47] ideas get solved. You know, we come up [42:49] with new ways of doing things. So, I'm [42:50] I'm just uh excited that uh for this [42:52] next phase of you know, never being a [42:54] better time to be working in a big [42:55] business, but also never being a better [42:57] time to be an entrepreneur. Um it's [42:59] going to be pretty wild. [43:00] >> Cambrian explosion of productivity, [43:02] check. All right, let's do a ching. [43:03] >> I'm I'm excited for that, too. Um [43:05] I mean, uh maybe I'll say it this way. [43:08] So, you mentioned evals earlier, right? [43:09] In some sense, the agents are doing the [43:11] work, so now what do the humans do? [43:12] Well, the humans have to verify and [43:16] think about what to do. And actually, [43:17] what I'm finding, what we're finding in [43:19] our teams, um by the way, I think I [43:20] learned the term AI vampire from someone [43:22] here. I'm not sure who, but um but [43:25] >> By Stevie. [43:25] >> Or yeah, but also by Stevie, yeah. But [43:28] uh [43:28] I you know, humans are actually getting [43:31] pushed to do more of the thinking. In [43:33] some sense, like people are busier than [43:34] ever, and it's actually harder than ever [43:37] because I can't just like spend 3 days [43:39] executing on the code. Like, I need to [43:41] actually think about what I want. And [43:42] that's actually harder. [43:44] Um and that's actually uh it's going to [43:46] be exciting. It's going to be fun. Like, [43:47] I think people will be pushed to sort of [43:48] the problem-solving and creativity uh [43:50] layers, and and that's going to be [43:51] really interesting. [43:52] >> Taste, judgment become far more [43:54] important and valuable. Well said. [43:56] Jessica? [43:57] >> Okay, you guys took all the good ones. [43:58] Um [43:59] >> [laughter] [43:59] >> so, uh [44:01] let me just let me just say that the [44:03] ceiling for an individual is like out of [44:06] sight. Like, I don't even I can't even [44:07] see it anymore. Like for example, um [44:10] I read the Gastown article created Claw [44:13] Town in a weekend. Like that is [44:16] something that I couldn't even imagine. [44:18] It just like fell in my lap. I couldn't [44:20] imagine this could happen to a single [44:22] person. Um so, I think I'm really [44:25] excited for a lot of like people from [44:28] uh a lot of people coming in through the [44:30] door. Like anyone can be a builder now. [44:32] Um and it's just [44:34] it's a really incredible time to be [44:36] alive and you get to build the next best [44:40] thing. [44:41] >> And you're helping a little bit with [44:42] your little support buddy to make that [44:44] future possible. [44:45] >> Mhm. [44:46] >> Uh all right, Ben. You get the last word [44:48] on this. [44:48] >> Uh what am I most excited about for the [44:51] future of AI and agents? [44:56] Well, I think that [44:59] what we've seen from the introduction of [45:02] agents when they when they first [45:05] >> landed here [45:06] >> is you could suddenly get all of your [45:09] questions answered really fast. You [45:11] could start to build products if you [45:13] didn't know how to code. And I think [45:15] that we're still at the very, very, very [45:17] beginning of what that's going to look [45:20] like. So, similar to the explosion of [45:23] productivity, what I'm most excited [45:25] about is the explosion of [45:29] uh positive uh impact on the world that [45:32] people can unleash with their [45:33] creativity. Right now, for example, I [45:36] don't see anybody talking to an agent [45:38] and the amount of lost opportunity [45:41] >> [laughter] [45:42] >> currently happening that uh [45:44] of the things everyone here has [45:46] something in your head that you want to [45:48] make. And I think the challenge for you [45:50] is to say [45:52] could I go do that in a weekend? Because [45:54] historically, you could not. And the [45:56] biggest thing you have to break through [45:58] as a human today in the world of AI is [46:01] to uh try. So, what I would say is go [46:05] home or go in the lobby or start doing [46:07] it right now and think of the craziest [46:09] thing that you wish you could do and [46:11] even if you don't think it's possible, [46:12] try to start doing it with an agent and [46:14] you are going to surprise yourself [46:17] because you probably can do it. [46:18] >> Well, I'm going to be looking forward to [46:20] that blockbuster movie your agent army [46:21] is going to create and and hopefully [46:23] it'll be just as much fun as you are [46:25] making us real. Let's give a big round [46:26] of applause for our wonderful panelists. [46:30] Hopefully, you all had a fascinating [46:33] discussion and learned and and going to [46:35] try out some of those things that you [46:37] that Matt is asking you to do. All [46:39] right. Thank you, everyone.