Structured Content is Foundational for AI Agent Operations

The AI agent revolution is not just about writing code faster; it's about fundamentally reshaping how we manage and operationalize information. This conversation with Simen Svale, co-founder and CTO of Sanity, reveals a critical, often overlooked, consequence: as AI agents become our primary interface with digital systems, the quality, structure, and accessibility of our content become paramount. The non-obvious implication is that the "headless" approach to content management, once a niche for developers, is now the foundational requirement for effective AI operations. Anyone building or managing digital products, especially those looking to leverage AI for content creation, analysis, or customer interaction, will gain a significant advantage by understanding this shift. Ignoring it means risking AI agents being fed bad data, leading to confusion, inefficiency, and potentially costly errors, mirroring the historical challenges of content management but amplified by the speed and scale of AI.

The Hidden Cost of "Easy" AI: Why Structured Content is the New Competitive Moat

The narrative around AI agents often focuses on their ability to write code, generate text, or automate tasks at unprecedented speeds. But this conversation with Simen Svale of Sanity cuts through the hype, revealing a more complex, and arguably more critical, layer: the foundational role of structured content. The immediate benefit of AI agents is undeniable, but the downstream consequences of feeding them unstructured, disorganized, or context-deficient data are starting to surface, creating a new battleground for competitive advantage.

The core problem, as Svale articulates, is not that AI agents are becoming unintelligent; quite the opposite. Their increasing sophistication means they are becoming incredibly adept at acting upon the information they are given. The danger lies in that information being flawed. Historically, managing content meant ensuring it was presentable on a website. The shift to "headless" CMSs, like Sanity, offered flexibility by decoupling content from presentation. This liberation, however, was primarily seen as a developer benefit, enabling multi-channel delivery. Now, with AI agents as a primary interface, this structured, data-driven approach is no longer just about websites; it's about providing reliable context for AI decision-making.

"The thing that isn't going away is giving your agents bad data--like if you give your your agents kind of weird fragmented data with like weird inside jokes or very contextually specific content to a specific page they will do odd things."

This highlights a crucial consequence: the "obvious" solution of simply dumping all your content into a repository for AI agents to access can backfire spectacularly. If that content is messy, inconsistent, or lacks proper referencing, the AI will act on that bad data. This is not a hypothetical future; it’s a present reality mirroring the early days of web content management. Companies once struggled with managing hundreds of disparate content systems, leading to synchronization nightmares. Now, Svale observes, this history is repeating itself with bespoke agent tools and context management systems. Teams are building their own systems, often requiring technical expertise like Git, to manage AI agent context, effectively creating a new barrier to entry for subject matter experts.

The real competitive advantage, therefore, lies not in having AI agents, but in having them operate on a foundation of well-modeled, accessible data. Sanity's approach, which Svale describes as a "content backend" and a "json database," is designed precisely for this. It provides the schema and structure that AI agents need to function reliably. When a business decision changes, like altering a pricing plan, a structured content system allows an agent to dynamically update all relevant instances across websites, documentation, and marketing materials. Without this structure, a human would have to manually find and update each piece of content, a process that is slow, error-prone, and impossible to scale with AI.

"The point is like if you can if you are working in a search a siloed or solo universe where you can just have a file and that's your kind of working matter and it's fine for you to just put that on the server then it's a very very small operation if you're any if you're if you're 10 people even you need to manage that coordinate that you need to stage those things."

The conversation around Lee Robinson's move to Cursor and his assertion about pulling content down into a repository sparked a valuable debate. While the immediate appeal of having all content as files in a repo for agent access is understandable, Svale's team pointed out that this often involves recreating the functionality of a CMS. The "slop-forking" approach, as it was termed, can lead to reinventing wheels that are already robustly engineered. The difficulty lies in designing effective "Multi-Cloud Protocols" (MCPs) -- essentially, protocols that define how agents interact with tools and services. Sanity's MCP, for instance, is designed to be simple and self-documenting, allowing agents to understand and utilize content management capabilities without requiring the user to be a technical expert. This abstraction is key. It allows non-technical users, or AI agents acting on their behalf, to manage complex content structures without needing to understand the underlying mechanics.

The distinction between "solved" and "actually improved" is critical here. A quick fix might solve an immediate problem, but it often creates downstream complexity. For example, a company might use AI to generate marketing copy, but if the underlying content isn't structured, the AI might produce inconsistent messaging across different platforms, requiring significant human oversight to correct. The "hard work" of content modeling, which might seem like an unnecessary upfront investment, pays off significantly by enabling AI agents to operate efficiently and reliably. This is where delayed payoffs create a lasting competitive advantage. Companies that invest in structured content now will find their AI agents are more effective, more reliable, and ultimately, more valuable. Those who don't will be left managing AI-generated chaos, constantly battling the downstream effects of poorly managed information.

The Unseen Power of Agentic Workflows

The discussion around MCPs and agentic workflows reveals a deeper systemic consequence: the shift from direct human interaction with interfaces to indirect interaction via AI agents. This necessitates a re-evaluation of how information is presented and managed. Tools like Pencil, a design application that functions as an MCP server, exemplify this. Users can instruct an AI agent to design interfaces, and the agent interacts with Pencil's tools and knowledge base. The output is a new file format, a "pen file," which is designed to be intuitively understood by agents, even without direct access to Pencil. This illustrates how specialized tools, when integrated into an agentic workflow, can abstract complexity and enable sophisticated tasks through natural language commands.

"The beauty of it of course is that you that they work everywhere it's such a simple dumb protocol so it's so it's implemented the user does not call these api endpoints no you don't you just have to know one like for for for http mcps the good ones you just need to know a url you just need to to put that in and then you do a login thing to become yourself and now the agent has that mcp."

This highlights the power of well-designed protocols. An agent doesn't need to understand the intricacies of Sanity's backend; it just needs to know how to interact with its MCP. This allows for a seamless flow of information and action. The conversation also touched upon Ingest, a workflow system for running long-running, durable processes. This is crucial for complex AI tasks, such as content analysis or web crawling, where processes might fail or require coordination. Ingest ensures these tasks are completed reliably, even if servers crash. It acts as an advanced queue and workflow driver, enabling agents to orchestrate complex operations without constant human supervision. The implication is clear: robust AI operations require not just intelligent agents but also reliable infrastructure to support their tasks.

The challenge of context windows -- the amount of information an AI model can process at once -- was also a recurring theme. Effectively summarizing vast amounts of data, like billions of documents or thousands of customer journeys, into a format usable by an agent is a significant hurdle. This compression problem is driving innovation in how information is stored and retrieved. The strategy of "offboarding" AI to deterministic flows, using classical automation or specialized agents for specific tasks, is emerging as a best practice. This means leveraging traditional software for predictable tasks (like full-text search in Sanity) and reserving AI for decision-making or complex reasoning. This hybrid approach optimizes for both efficiency and cost, ensuring that AI is used where it adds the most value, rather than being applied to every problem.

Finally, the conversation explored the "model whisperer" phenomenon -- the need to craft prompts and instructions carefully to elicit the desired behavior from AI agents. The inherent "people-pleasing" nature of current models means they often agree with flawed logic. Developing strategies to encourage critical feedback, such as using a dedicated "devil's advocate" agent, is becoming essential. This highlights that even as AI becomes more capable, human oversight and intelligent design of AI interactions remain critical for achieving optimal outcomes. The future of AI-powered work isn't just about automation; it's about sophisticated collaboration between humans, structured data, and intelligent agents, all orchestrated through well-designed systems.

• Immediate Action: Implement a structured content modeling strategy for all new content initiatives.
• Immediate Action: Audit existing content repositories for consistency, completeness, and clear referencing.
• Immediate Action: Explore and integrate Multi-Cloud Protocols (MCPs) for key AI agent interactions, starting with content retrieval.
• Longer-Term Investment (6-12 months): Develop or adopt a centralized content backend system (like Sanity) to serve as the single source of truth for all digital content.
• Longer-Term Investment (12-18 months): Investigate workflow automation tools (like Ingest) to manage complex, long-running AI agent tasks, ensuring durability and reliability.
• Discomfort Now for Advantage Later: Dedicate resources to training subject matter experts on how to effectively manage content within a structured system, even if it requires a learning curve.
• Discomfort Now for Advantage Later: Proactively design prompts and interaction patterns for AI agents that encourage critical feedback and discourage blind agreement, even if it feels counter-intuitive.