Software in the Age of Agents | The a16z Show

这期a16z播客邀请了Seema Amble和Steven Sinofsky，他们深入探讨了“无头软件”及其对企业软件的深远影响。对话深入探讨了定义粘性软件的要素、智能代理和人工智能如何颠覆传统范式，以及为初创企业带来的巨大机遇。 **理解无头软件** Seema Amble 将“无头软件”定义为并非一个新概念，而是一个正获得广泛关注的理念。虽然 Salesforce 的“无头360”发布主要是一项营销举措，但它承认了一个关键转变：传统软件围绕人机交互和用户界面（UI）构建，但在一个“代理世界”中，用户界面变得不那么重要了。真正的价值现在在于底层数据和逻辑。智能代理，无论是复杂的API还是简单的聊天机器人（例如与Salesforce交互的Slack机器人），都可以直接访问记录系统，绕过传统界面。Steven Sinofsky 诙谐地指出，“代理”有时可能只是一个长期运行程序的“换壳”，但他将代理交互分为三类：“查找”（简单查询）、“执行操作”（需要凭据并导致系统更改）和“分析”（涉及多个系统，如果未经核实则容易产生幻觉）。 **传统企业软件的粘性** 历史上，软件的粘性源于人机交互。用户界面培养了肌肉记忆，未成文的标准操作程序变得根深蒂固，而部门间的依赖性巩固了产品的地位。Steven Sinofsky 补充说，向客户收款的行为是最终的粘性所在，这使得更换软件变得极其困难。公司因特定的法规（如 HIPAA）、管理需求（如入职流程），甚至是一些神秘的、偶然的功能（如微软Outlook的日历委派功能）而与软件深度交织，这些功能变得不可或缺。软件“渗透”到组织核心业务中的这种现象产生了巨大的惯性。 **SAP：粘性的终极典范** 专家们强调 SAP 是粘性软件的典型范例。它不仅仅是一个数据库；它固化了整个公司的运营逻辑，通常是出于法律或监管原因。对于沃尔玛或大型汽车制造商等大型企业来说，替换 SAP 几乎是不可能的，因为它们的业务流程正是由 SAP 的定制化所定义的。初创企业常常低估这种复杂性，错误地认为大型企业的需求只是简单的电子表格替代品。拉里·埃里森（Larry Ellison）过去关于公司只需要80%解决方案的咆哮，很大程度上被驳回了，因为企业的差异化在于它们如何处理剩下的20%——即嵌入在其ERP中的独特定制和流程。 **人工智能的作用：增强，而非总是取代** 虽然人工智能和智能代理不会简单地取代 SAP 等系统，但它们提供了显著的增强机会。人工智能可以提取和合成信息，实现自然语言查询和定制报告，而无需操作复杂的界面。这关乎让现有、数据丰富的系统*更易用*、更易访问。 然而，一个主要挑战在于捕捉“上下文”——即人们头脑中那些细微的例外情况、政策和不成文的规则。智能代理需要这些上下文才能超越基本数据检索，并有效地“行动”。亚马逊的客户服务方法，优先考虑客户满意度，然后利用数据改进内部流程，为通过自动化处理异常情况提供了一个有趣的范例。 **挑战与误解** 自动化处理“长尾”异常情况异常困难。此外，当多个智能代理读写系统时，权限管理、多代理交互以及维护单一真相来源等问题变得复杂。一个显著的误解是，自动化仅仅是减少了现有工作。相反，生产力提升往往导致*新*任务和场景的产生。例如，费用报销从基本的会计核算发展到复杂的商务旅行分析。企业软件供应商也抵制去中介化；他们不希望被降级为“笨数据库”，并将继续开发功能，即使这些功能并非总是最好的。 **初创企业的机遇** 初创企业有三条主要路径： 1. **依附现有巨头：** 将智能代理与Salesforce等现有系统集成，尽管现有巨头可能抵制成为单纯的后端。 2. **自行替换：** 对于复杂的企业来说极其困难，堪比“心脏直视手术”。 3. **人工智能软件初创企业：** 最有前景。这些公司可以*与*现有系统协同工作，增强可见性，收集新数据（例如，来自语音代理、文档摄取），并提供可操作的情报（例如，潜在客户优先级排序、客户流失预测）。它们发现现有巨头*没有*做的事情。 最大的机会是在现有参与者*之间*运作。初创企业无需正面竞争，可以通过利用人工智能来连接组织内以前未能有效沟通的职能部门（例如，设计与产品、IT与财务），从而创建新的类别。这类似于网络，尽管最初在某些方面不如客户端-服务器架构，但它提供了一种全新的范式。 最后，Steven Sinofsky 指出，虽然企业软件中很难实现外部网络效应，但内部网络效应（即某个工具能让个人工作变得更好，并在组织内部像早期的 Excel 一样病毒式传播）是人工智能驱动工具的黄金机遇。

This a16z podcast features Seema Amble and Steven Sinofsky discussing "headless software" and its profound implications for enterprise software. The conversation delves into what defines sticky software, how agents and AI are disrupting traditional paradigms, and the vast opportunities emerging for startups. **Understanding Headless Software** Seema Amble defines "headless software" not as a new concept, but one gaining significant traction. While Salesforce's "headless 360" announcement was largely a marketing initiative, it acknowledged a crucial shift: traditional software was built around human interaction and UIs, but in an "agentic world," the UI becomes less relevant. The true value now lies in the underlying data and logic. Agents, whether sophisticated APIs or simple chatbots (like Slack bots interacting with Salesforce), access systems of record directly, bypassing traditional interfaces. Steven Sinofsky playfully notes that "agent" can sometimes be a rebrand for a long-running program, but distinguishes agent interaction into three categories: "Look Up" (simple queries), "Do Something" (requiring credentials and causing system changes), and "Analyze" (involving multiple systems and prone to hallucination if not verified). **The Stickiness of Traditional Enterprise Software** Historically, software stickiness stemmed from human interaction. UIs fostered muscle memory, undocumented standard operating procedures became ingrained, and inter-departmental dependencies solidified a product's position. Steven Sinofsky adds that the act of collecting money from a customer is the ultimate stickiness, making it incredibly hard to switch. Companies become deeply intertwined with software due to specific regulations (HIPAA), administrative needs (onboarding), or even arcane, accidental features (like Microsoft Outlook's calendar delegation) that become indispensable. This "seepage" of software into an organization's core operations creates immense inertia. **SAP: The Ultimate Example of Stickiness** The panelists highlight SAP as the quintessential example of sticky software. It's not just a database; it codifies an entire company's operational logic, often for legal or regulatory reasons. Replacing SAP is practically impossible for large enterprises like Walmart or major auto manufacturers, as their very business processes are defined by its customizations. Startups often underestimate this complexity, mistaking a large enterprise's needs for simple spreadsheet replacements. Larry Ellison's past rant about companies only needing 80% solutions was largely dismissed because businesses are differentiated by how they handle the remaining 20% – the unique customizations and processes embedded within their ERPs. **AI's Role: Enhancing, Not Always Replacing** While AI and agents won't simply replace systems like SAP, they offer significant opportunities for enhancement. AI can extract and synthesize information, enabling natural language queries and customized reports without navigating complex UIs. It's about making existing, data-rich systems *more usable* and accessible. However, a major challenge lies in capturing "context" – the nuanced exceptions, policies, and unwritten rules that reside in people's heads. Agents need this context to move beyond basic data retrieval and effectively "act." Amazon's customer service approach, which prioritizes customer satisfaction and then uses data to improve internal processes, serves as an interesting model for handling exceptions through automation. **Challenges and Misconceptions** Automating the "long tail" of exceptions is incredibly hard. Furthermore, issues like permissioning, multi-agent interaction, and maintaining a single source of truth become complex when multiple agents are reading and writing to systems. A significant misconception is that automation merely reduces existing work. Instead, productivity gains often lead to the creation of *new* tasks and scenarios. For example, expense reporting evolved from basic accounting to sophisticated business travel analysis. Enterprise software vendors also resist disintermediation; they don't want to be relegated to "dumb databases" and will continue to build features, even if they're not always the best. **Opportunities for Startups** Startups have three main paths: 1. **Building on Incumbents:** Integrating agents with existing systems like Salesforce, though incumbents may resist being mere backends. 2. **DIY Replacement:** Extremely difficult for complex enterprises, akin to "open heart surgery." 3. **AI Software Startups:** The most promising. These companies can work *alongside* existing systems, enhancing visibility, collecting new data (e.g., from voice agents, document ingestion), and providing actionable intelligence (e.g., lead prioritization, churn prediction). They identify what incumbents *aren't* doing. The biggest opportunity is to operate *between* established players. Instead of competing head-on, startups can create new categories by leveraging AI to bridge functions within an organization that previously didn't communicate effectively (e.g., design and product, IT and finance). This is analogous to how the web, despite being inferior to client-server in some ways initially, offered a completely new paradigm. Finally, Steven Sinofsky notes that while external network effects are hard in enterprise software, internal network effects (where a tool makes individual jobs better and spreads virally within an organization, much like early Excel) are a golden opportunity for AI-driven tools.