Big Tech Backs New AI Agent Discovery Standard While Enterprise Leaders Warn Against Handing Over the Keys

A New Standard for AI Agents to Find Their Own Tools

On June 18, 2026, Google and Microsoft published blog posts announcing a new open specification called Agentic Resource Discovery (ARD). The backing coalition includes NVIDIA, Salesforce, Cisco, Snowflake, Databricks, GitHub, GoDaddy, Hugging Face, and ServiceNow, according to ZDNET's David Gewirtz.

The goal: let AI agents discover, evaluate, and use tools, skills, and other agents at runtime without a human having to configure each connection manually.

The analogy Gewirtz uses is useful. Anthropic's Model Context Protocol (MCP), introduced in 2024, standardized how AI systems communicate with servers. MCP made the apps possible. ARD is meant to be the app store, a discovery layer that lets agents find what they need on the fly.

Noticeably absent from the ARD coalition: OpenAI and Anthropic. ZDNET does not explain why, and neither Google nor Microsoft addressed the omission in their announcements. That's a significant gap worth watching, given that OpenAI and Anthropic together power a large share of enterprise AI deployments today.

The Security Problem Nobody Is Solving Yet

Gewirtz raises a concern that the announcement glosses over: a discovery layer for AI agents is also a new attack surface. If agents can search for and connect to external tools at runtime, they can also be pointed at malicious ones.

The ARD specification is described as covering publishing, discovering, and verifying AI capabilities. But what verification means in practice, and who enforces it, is not detailed in the sources available as of June 18, 2026. Whether ARD becomes infrastructure or a liability depends on those details.

The strongest counterargument to that security concern is real. Without a standard like ARD, enterprises are already duct-taping together agent workflows with no consistent governance at all. A well-designed open standard, if the verification layer holds up, could actually reduce the chaos. The concern isn't that a standard is being created. It's whether this consortium can deliver on the verification promise before deployment outruns the spec.

Enterprise Reality: Slow Down to Go Fast

While the ARD announcement envisions agents operating with increasing autonomy, two senior enterprise executives speaking at a conference hosted by consultancy Section offered a ground-level counterweight.

Scott Likens, global chief AI engineer at PwC, put it directly: "Stop being the human in the loop. The human is the loop." His point is not that AI should run itself. It's the opposite. Human oversight shouldn't be a checkpoint inside an automated process. It should be the architecture the process is built around.

Lasherelle Morgan, senior vice president of AI innovation and acceleration at NBCUniversal, said the right starting point is the pain, not the technology. "Don't just bring in an AI tool. Ask, 'what are you struggling with?' 'What are you spending five hours of your day on?'" she said, according to ZDNET's Joe McKendrick.

PwC runs AI experiments in one-day or five-day cycles, according to Likens. That cadence is fast, but it's structured. It starts with a specific business problem, not a capability demonstration.

The Cost-Savings Trap

Likens also flagged a mindset problem spreading through enterprise AI: the fixation on token costs and narrow efficiency gains. "All this talk of tokens just started a couple of months ago, and now all of a sudden there is a cost focus with AI," he said. "That's the wrong way to look at it."

The argument is that companies squeezing for 2–3% cost savings are undershooting what AI can actually change. Experimentation, not optimization, is where the real value gets found. That framing cuts against the current enterprise trend of treating AI primarily as a headcount-reduction tool.

Two Competing Impulses, One Industry

The ARD announcement and the enterprise warnings aren't contradictory. They're describing different layers of the same problem. ARD is infrastructure. Human-centered deployment is governance. Neither works without the other.

But the timing matters. A discovery standard that lets AI agents autonomously find and connect to new capabilities is being announced on the same day that senior enterprise leaders are warning companies not to let AI run without direct human accountability. The infrastructure is being built faster than the governance frameworks companies need to use it safely.

The concrete question ARD leaves open: who audits verified capabilities in the registry, and what happens when a verified tool behaves badly after deployment? Neither Google's nor Microsoft's announcement, as reported by ZDNET, answers that.