AI standards and protocols: the open specs that wire the agentic compute stack

What it is

This beat tracks the formal specifications that determine how AI systems interconnect and the governance bodies that write them. Two distinct layers matter. The software layer covers how agents call external tools and coordinate with other agents: the Model Context Protocol (MCP) and Agent2Agent (A2A) are the key interfaces. The hardware layer covers how AI accelerators link inside training clusters: the Ultra Accelerator Link (UALink) standard governs chip-to-chip scale-up within a pod, and the Ultra Ethernet Consortium's specification covers scale-out between pods. A world-news reader tracks this because whoever writes these specs sets the interoperability rules for the agentic economy, from which vendors can sell into enterprise AI budgets to how governments regulate cross-border AI system deployments.

History

International AI standards-setting took institutional form in 2017, when ISO and the International Electrotechnical Commission established ISO/IEC JTC1/SC42, the joint committee that has since published 45 standards covering AI management systems, trustworthiness, and governance. The US National Institute of Standards and Technology (NIST) published its AI Risk Management Framework in January 2023, a voluntary vocabulary that influenced subsequent EU and UK frameworks.

The protocol-layer inflection came in November 2024, when US company Anthropic open-sourced the Model Context Protocol, a vendor-neutral client-server interface that replaced bespoke per-framework tool wiring. MCP reached roughly 97 million cumulative downloads by mid-2026. In December 2025, Anthropic donated MCP to the newly formed Agentic AI Foundation (AAIF), a directed fund of the Linux Foundation co-founded with Block and OpenAI, governed by eight Platinum members including AWS, Google, and Microsoft.

On the hardware side, the UALink Consortium published four specifications in April 2026 for chip-to-chip scale-up interconnect. The Ultra Ethernet Consortium released a 560-page Specification 1.0 bringing InfiniBand-class RDMA and congestion control to commodity Ethernet gear.

Current state

As of July 2026, the software protocol stack has converged on a two-layer architecture: MCP for vertical agent-to-tool calls and A2A for horizontal agent-to-agent delegation. The 28 July 2026 MCP specification makes the protocol stateless via six specification enhancement proposals and adds sandboxed MCP Apps. A2A reached v1.2 with cryptographically signed agent cards, in production at Microsoft, AWS, Salesforce, SAP, and ServiceNow across more than 150 organisations. A joint MCP/A2A specification targeting shared authorization is due in Q3 2026.

On hardware, UALink's four April 2026 specifications cover in-network compute, a 200G physical layer, manageability, and UCIe 3.0-compliant chiplet integration, competing with Nvidia's NVLink in the scale-up pod market. Ultra Ethernet's 560-page Specification 1.0 brings RDMA and congestion control to commodity Ethernet at InfiniBand-class performance, positioning it as the cost-led alternative for scale-out clusters.

NIST's Center for AI Standards and Innovation formally launched its AI Agent Standards Initiative in February 2026, covering three pillars: industry standards facilitation, open-source protocol support, and agent authentication research. A public comment period closed in March 2026.

Relationships

The three related nodes map the live edges of this beat. Agent protocols harden: MCP goes stateless, A2A passes 150 orgs covers the software protocol layer: MCP's stateless July 2026 specification and A2A's signed agent cards define how agents call tools and delegate tasks across vendors. UALink, 4개 규격 공개하며 NVLink에 맞선 개방형 대안 공식화 covers the chip-level scale-up fabric: UALink's four specifications target the 1,024-accelerator pod market that Nvidia's NVLink currently dominates. Ultra Ethernet 1.0 turns commodity Ethernet into an AI back-end fabric covers the rack-level scale-out fabric: Ultra Ethernet's RDMA-capable specification is the industry's cost-led alternative to InfiniBand for million-GPU back-end networks. Governance sits above all three: AAIF steers the agent protocol roadmaps, NIST frames the US federal baseline, and ISO/IEC JTC1/SC42 covers the international layer with 45 published AI standards.

What to watch

The Q3 2026 joint MCP/A2A specification, where shared authorization remains the missing link for cross-vendor multi-agent enterprise deployments. Whether NIST's AI Agent Standards Initiative produces formal guidelines by end-2026 and whether those guidelines align with EU AI Act implementing measures. First UALink silicon from AMD, Intel, and Astera Labs, expected late 2026, and whether it interoperates across vendors in practice. Geopolitical competition in standards bodies: the US and China are each pressing AI frameworks at ISO/IEC JTC1/SC42 and the ITU, and which proposals carry majority votes will shape the international interoperability and security baseline.