Salesforce Knows Einstein Is Broken

Salesforce Research just dropped OpenAgents, an open platform for language agents, and it's good — genuinely good, not press-release good — which is worth noting precisely because of who published it and why they might be motivated to work on this.

Einstein is limited. Disconnected. Anyone who has actually tried to get it to do something interesting with Data Cloud knows this. You have this massive unified data layer sitting there and the AI assistant that's supposed to sit on top of it can't really reach it in any meaningful way. That's the context you need to read this paper with.

What OpenAgents is building is the architecture to close that gap: user asks a simple question, the agent generates a hidden, complex intermediate answer — queries, plans, tool calls, whatever the job requires — and then presents a clean, grounded result back up top. Simple in, simple out, with all the ugly machinery in the middle invisible to the person who just wants to know something true about their data.

They've got three agents. A Data Agent for code-assisted analytics. A Web Agent for autonomous browsing. And the Plugins Agent, which is the interesting one — 200+ everyday tools, real plugins, the kind of thing that makes this feel less like a research demo and more like something someone might actually use. They show it off with a GetYourGuide integration, planning a trip to Toronto, which is the least interesting possible showcase for something that could, in the right hands, be pointed at a Salesforce org and told to figure out why pipeline is down 23% in the northeast region.

The plugins architecture is the part worth studying. It's not just "here are tools the LLM can call." It's a coordination layer — the agent learns what's available, decides what's relevant, calls what it needs, stitches the outputs into something coherent. That's the machinery Einstein is missing. Not the model. Not even the data access. The coordination layer that makes a question feel answered rather than processed.

One more thing. If you throw a Mixture of Experts model at the building step — the part where the agent figures out what to do — the elaborate multi-step walkthrough they show in the paper starts to look like unnecessary overhead. A sufficiently expert router probably collapses several of those steps. Nobody in the paper says this. But it's there.

Salesforce Research publishing this openly, right now, is not a coincidence. They know what they have. They know what Einstein isn't. This is the work.