The rails, not the models — the bet enterprise AI just started making

THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION

DEEP DIVE · COMPANION TO ISSUE #143

The rails, not the models — the bet enterprise AI just started making

Eleven enterprise giants published a standard for how AI agents find their tools. The two labs with the best models weren't on the list. Here's what that does — and doesn't — prove.

BY R. LAURITSEN · 15 JULY 2026 · 8 MIN READ

Read the supporter list first. Google. Microsoft. Salesforce. Cisco, Databricks, GitHub, GoDaddy, Hugging Face, NVIDIA, ServiceNow, Snowflake. Eleven of the most powerful names in enterprise software, backing one open specification for how AI agents discover the tools they can use. Now notice who isn't there: OpenAI and Anthropic. What this piece is actually about is what that absence means — because it could mean a great deal, or almost nothing, and most of the coverage this week hasn't bothered to tell the difference.

Last week this column argued the labs had started paying to shape the regulatory gate — buying influence over the rules that decide who ships a frontier model. That's still true. But this week a second contest opened, quieter and arguably more consequential, and the labs weren't the ones building it. This one isn't about permission to release a model. It's about whether, once released, that model can plug into the software where actual work happens.

What ARD actually is — and why the name undersells it

The mechanism (status: shipped, Apache-2.0). Agentic Resource Discovery lets an organisation publish an ai-catalog.json file — a manifest of which tools, data sources and other agents live inside its systems. An agent handling a procurement request can then look up the approval system, the budget tool and the vendor database on its own, without an engineer hard-wiring each connection in advance. It's a phone book for agents. Dull-sounding. Foundational.

The layer it sits in (status: contested). ARD isn't alone. It's designed to complement Anthropic's Model Context Protocol — the standard for how an agent talks to a single tool — and Google's Agent2Agent protocol, which handles how agents talk to each other and is now running in 150 organisations in production, not pilot. Stack them and you get the plumbing of the agentic enterprise: MCP for the tool connection, A2A for the agent conversation, ARD for the discovery. Google co-authors two of the three. One honest caveat: in practice these layers overlap and the boundaries are fuzzier than any diagram — treat the three-part split as a mental model, not a clean spec.

The tell (status: interpretation). MCP is Anthropic's. That matters — it's the one piece of this stack a frontier lab owns outright, and it's genuinely everywhere. But MCP is the narrowest layer: the wire between one agent and one tool. The layers above it — who's discoverable, who orchestrates whom — are being defined by the companies that also happen to own the software the agents plug into. Owning the wire is not the same as owning the map.

Why the best model may no longer be enough

Be precise about which moat is under discussion. Not model capability — that still matters, and a smarter model is still worth more than a dumber one. The moat in question is enterprise distribution and integration: whether your model can reach the systems where work actually happens. For three years the assumption was that the best model wins, and it held while AI was a chat box — a person picked the smartest assistant and typed at it. But the enterprise isn't buying chat boxes any more. It's buying agents that run multi-step work across a dozen systems, and an agent is only as useful as the tools it can reach.

That flips the economics. When the constraint moves from “how smart is the model” to “what can the model actually touch,” the advantage moves to whoever controls the touching. Google made the argument out loud at Cloud Next: it owns the chip (Ironwood TPUs), the model (Gemini), the runtime and the distribution channel — three billion users across Workspace. Its pitch to enterprises was a jab at exactly the two absent labs: rivals “hand you the pieces, not the platform.” Harsh. Also basically correct as a description of where OpenAI and Anthropic sit — extraordinary pieces, no platform underneath them.

Here's the cascade that should worry them. If the best model is a guest on Google, Microsoft and Salesforce's rails, then those rails can set the terms — pricing, priority, default routing. A model that's fractionally better but has to be manually wired into every workflow loses to a model that's fractionally worse but is already discoverable, already orchestrated, already sitting in the catalog the agent reads first. At the margin, convenience can beat quality — and the enterprise is nothing but margins.

The counter-arguments, taken seriously

“ARD is open — anyone can join.” This is the strongest objection, it deserves the most weight, and I can't fully refute it. Apache-2.0 means OpenAI could implement ARD next week and Anthropic the week after; nobody is technically locked out. Standards do tend to get shaped by whoever shows up early enough to set defaults and seed the ecosystem — but whether that dynamic actually plays out with ARD is unknown today. The spec is days old. No default behaviour has been set. If a frontier lab formally backs it within the quarter, this whole thesis weakens considerably — and I'd rather flag that up front than bury it under a metaphor.

“MCP already won the important layer.” Also fair. MCP has tens of thousands of servers and tens of millions of monthly SDK downloads — it's the closest thing to a universal tool-connection standard going, and it's Anthropic's. If discovery and orchestration end up mattering less than the raw tool connection, Anthropic is sitting pretty. The bet ARD's backers are making is the opposite: that in a world of thousands of agents, finding and coordinating them is the hard part, and the wire is the easy part.

“The labs don't need the enterprise stack — they'll go direct.” This is what the forward-deployed engineering armies are for. OpenAI's roughly $4bn deployment venture, Anthropic's $1.5bn, AWS's new $1bn org — these are the labs paying humans to stand where the rails would be, embedding engineers to hand-build what a standard would otherwise automate. It works. It also doesn't scale the way a protocol does.

The bet, stated plainly

By the end of Q4 2026, at least one frontier lab formally backs ARD — a public commitment or a shipped reference implementation, not a passing compatibility mention — or announces a competing discovery standard of its own. Setting the bar before the deadline is the point: it stops me claiming a vague announcement as vindication later. Both OpenAI and Anthropic move fast when they sense a structural threat, and their deployment ventures are consistent with that. When the announcement comes, it'll be framed as “embracing open interoperability” — which might be the whole truth (open compatibility is cheap and sensible) or might be the moment reach started mattering as much as reasoning. The forecast is falsifiable either way; the motive is the reader's to judge.

Last week the question was “who profits from the permission list existing?” This week's is quieter and lands closer to home: when your agent goes looking for its tools, whose map is it reading — and who drew it? Reply to this week's issue if you think I've called it wrong. The sharpest counter-argument gets printed next week.

iPrompt

PUBLISHED BY FRONTWAVE MEDIA LTD · LIMASSOL, CYPRUS · IPROMPT.COM