Software in the Age of Agents | The a16z Show

This discussion explores the profound shift in enterprise software as AI agents become primary users, moving away from human-centric UIs. The conversation begins by unpacking the concept of “headless software,” exemplified by Salesforce’s “Headless 360” announcement. While Salesforce’s move was largely a re-branding of existing APIs, it signifies an acknowledgement of a fundamental shift: the value of software increasingly resides in its underlying data and logic, rather than the graphical user interface. Agents, like Slackbots experiencing a 300% increase in usage, are accessing information via APIs, making the UI less relevant for many interactions.

Agents, APIs, and Definitional Challenges

Steven Sinofsky highlights the “definitional hell” surrounding new terms like “agent,” humorously describing it as a program that takes a very long time to run. He categorizes agent actions into three types:

• Lookup: Lightweight, like retrieving information.
• Do Something: Involves impersonating users and credentials, raising complex enterprise software issues.
• Analyze: More complex, involving multiple systems, iterations, and potential for hallucination, requiring verification mechanisms. This type is particularly well-suited for agents due to its time-unbounded nature.

The Enduring Stickiness of Enterprise Software

Seema Amble and Steven Sinofsky delve into what makes enterprise software incredibly sticky:

• Human-centric design: Historically, UIs and workflows were built around human interaction, muscle memory, and standard operating procedures (SOPs).
• Ecosystem integration: CRMs like Salesforce become sticky due to widespread user adoption, reliance by finance and marketing, and external dependencies.
• Single source of truth: Critical for compliance (e.g., payroll, ERP) and operational consistency.
• Financial commitment: Software that collects money from customers (e.g., Stripe) creates significant inertia.
• Codification of business rules: Systems like SAP don’t just store data; they embed and enforce an organization’s unique operational processes, regulatory requirements, and even corporate culture. Displacing them would effectively dismantle the company itself.
• Unexpected arcane features: Often, software becomes indispensable due to niche functionalities (like Outlook’s delegate access) that were not intentionally designed for stickiness but became critical for specific users or industries.

The “Death of Software” is Overblown

The notion of software being easily replaced is strongly refuted. Steven emphasizes that systems like SAP are as sticky as the insurance software written 50-75 years ago, which codified regulatory bodies’ rules. He criticizes the misconception that a PostgreSQL database and APIs can replace SAP, stressing that the captured logic is far more important than just the data. Larry Ellison’s past rant about customizing software and advocating for an “80% solution” is presented as a prime example of underestimating the sophistication and customization required by large enterprises. Companies like Goldman Sachs leveraging Excel, or auto manufacturers relying on SAP for their entire operational blueprint, illustrate this deep integration.

The True Game: Exception Handling

The speakers argue that exception handling is the entire game in enterprise software. While agents can automate mundane tasks, the true complexity lies in dealing with the long tail of exceptions and edge cases that define how a business operates. These exceptions are often unwritten or exist in human knowledge, not explicitly captured in software fields. Amazon’s customer service approach of

• Tags:
• AI Agents
• Feature
• Workflow Automation
• SaaS
• Enterprise Software
• APIs
• Middleware
• Network Effects
• SAP
• Salesforce
• Headless Software
• Software Stickiness
• Exception Handling
• Startup Opportunities