ARD ships. AI agents can now find your tools. That's not the hard part.

On June 17, Google published the Agentic Resource Discovery specification — a draft open standard co-developed with Microsoft, Salesforce, GitHub, Hugging Face, NVIDIA, Snowflake, Cisco, Databricks, GoDaddy, and ServiceNow. Eleven companies. Apache 2.0 licence. Reference implementations shipped the same day.

The problem ARD solves is real. Right now, an AI agent can only use a tool that someone manually wired to it ahead of time. Microsoft's Ramanathan Guha put it plainly: "AI can only use what it's been explicitly wired to use. Everything else may as well not even exist." ARD fixes that by letting organisations publish an ai-catalog.json file on their own domain. Agents query a federated registry, find the right capability by plain-language intent, verify it, and connect — without a pre-built integration.

GitHub shipped a reference implementation the same day: Agent Finder for Copilot, which lets developers describe a task in natural language and get ranked matches from a curated catalog. Hugging Face released a Discover Tool for searching across skills and MCP servers. Google's Agent Registry in the Gemini Enterprise Agent Platform is adding native ARD support in the coming months.

For anyone building agentic marketing workflows, this is genuinely useful. The integration tax — the hours spent writing custom connectors so one agent can reach one API — is one of the most underappreciated friction points in the stack. I've felt it building workflows in n8n: you spend more time wiring tools together than you do on the actual logic. ARD is a plausible path to reducing that.

But there's a gap in the coverage I've read this week, and it matters for B2B marketers specifically.

ARD solves the runtime discovery problem. It does not touch the training-data problem. And for most B2B SaaS tools, the training-data problem is the more urgent one.

Here's how it actually works: when an agent starts a session, it doesn't begin with a blank slate. It begins with everything it absorbed during pretraining. Research published by Synscribe this week puts agents at roughly a five-month stale training prior in June 2026 — confidently operating on a January 2026 worldview. If an agent already "knows" your product from training — even with outdated, incorrect, or competitor-favourable data — it won't fire an ARD registry query. It answers from memory.

Publishing an ai-catalog.json won't fix that. Nothing in ARD touches your training presence.

The practical consequence: incumbents with strong training representation are being confidently recommended right now, even when newer or better alternatives exist. In workflow automation, the DerivateX AI Visibility benchmark from earlier this year showed Zapier scoring 63 and Make scoring 40 — despite Make appearing on all four platforms tested and Zapier being absent from Claude entirely. Mention frequency in training data doesn't track cleanly with product quality or even current market share. It tracks with how much was written about a tool, on authoritative sources, before the model's cutoff.

I use Make regularly. Seeing it score 23 points below Zapier in AI presence, when it's the more capable tool for most of what I'm building, is a concrete example of the gap ARD alone won't close.

None of this is an argument against ARD. The spec is clearly the right infrastructure to build. Dynamic discovery at runtime will matter enormously once agentic workflows become the default way marketing stacks operate — and that transition is moving faster than most teams have planned for. ARD is also explicitly designed to complement MCP and Agent2Agent, not replace them — so it slots into the stack rather than restarting the integration work that's already been done.

The thing worth separating out: ARD solves discovery for agents that know they don't know. The harder problem is agents that are confidently wrong about what they know — and that problem lives in training data, structured content, and AI search presence, not in catalog manifests.

If you're a B2B marketer thinking about your tool's visibility in an agentic world, the ARD spec is worth watching. Adding an ai-catalog.json to your domain is probably a 30-minute job once the registries are indexed. But the real work is upstream: being cited in the sources models train on, having structured, authoritative content that survives a training cutoff, and showing up consistently when AI platforms encounter buyer-intent queries about your category.

Runtime discovery is table stakes. Training presence is the moat.