AI Agents Need Context. Salesforce Just Bought It.

I’ve been sitting with the Salesforce/Contentful news since the acquisition was announced because I didn’t want to rush out a hot take. 🔥

The obvious read is straightforward: Salesforce wanted a structured content platform to support Agentforce, its AI agent platform, and Contentful had one.

That makes sense. But after sitting with it for a day, and after reading an insightful post by Michael Andrews, I think there’s a more interesting question underneath the deal.

Both Salesforce and Contentful have spent years selling a headless, API-first future. Andrews raised the question that I think many people skipped over: in that future, who manages the traffic? Who decides what an AI agent can access, what it’s allowed to do, and who can audit what happened after the fact?

That’s the conversation I think this acquisition is really about.

Since we started building Directus we’ve believed that content was never a separate category of software. Customer records, product catalogs, editorial content, documentation, operational data… they’re all variations of the same problem: structured information that needs to be governed, queried, and delivered to the right consumer at the right time.

The teams managing that information may have different job titles and different interfaces, but underneath they’re solving the same problem.

Salesforce didn’t just buy a content management system… it bought a major source of context for the AI agents it wants to power. Agentforce only becomes valuable when it can operate on more than CRM records. Product data, knowledge bases, documentation, support content, and operational systems all become part of an agent’s working context. The more useful the agent becomes, the more important that context becomes.

That’s what makes this acquisition interesting.

Before you deploy AI agents at scale, you need clear answers to some very practical questions. Can the agent access this information? Should it be allowed to? What is it permitted to say? To whom? Under what circumstances? And can we see what happened afterward?

Most organizations are still figuring all that out.

Connecting AI to information is rapidly becoming the easy part... but governing that access across hundreds or thousands of agents operating against production systems is much harder. That’s why we shipped a native MCP server in Directus. AI tools can query your data using natural language, but they operate through the same role-based access controls as human users. Claude, Gemini, or whatever comes next only sees the information its assigned role is permitted to access.

No carte blanche. No bypasses. No special treatment for the agent.

We built it that way because we believed governance would become more important (not less) as AI adoption accelerated.

Contentful deserves enormous credit for helping make composable architecture mainstream. API-first, best-of-breed, decoupled systems… they helped define that movement, and much of the industry followed (including us).

What’s interesting is that the company most associated with composability is now becoming part of a much larger platform. I’m not suggesting that’s the wrong move. Running composable environments is expensive. Integrations require maintenance. Every additional system introduces operational overhead. So for many organizations, consolidation is a decent trade.

The question is whether the tradeoffs are still worth it as the AI ecosystem evolves.

Anthropic, Google, OpenAI, and others are moving at extraordinary speed. Models improve constantly. New protocols emerge. New agent frameworks appear every few months. For organizations already deeply invested in Salesforce, tighter integration may remove meaningful friction. But for organizations building complex products, internal platforms, or systems that need flexibility across vendors and technologies, this might be a big step backwards.

The model itself is increasingly becoming the least durable part of the stack. What matters more is the context beneath it. The information available to the agent. The permissions governing access. The auditability of its actions. The ability to evolve as the ecosystem changes.

Salesforce’s bet is that owning the content layer strengthens Agentforce. On that point, they might be right. But the bigger problem isn’t just giving AI access to information... it’s governing that access responsibly and understanding what happens once those agents are operating in production.

That’s the problem I think the industry is only beginning to see and struggle with. And it’s the big problem that will still be there after today’s models are replaced by tomorrow’s.