ANALYSIS Enterprise AI is splitting into two distinct architectural layers: the agents that do work and the control planes that govern them. Across protocol foundations, startup launches, bank engineering teams, and survey data, the evidence points to a single structural shift — orchestration, context, and cost control are being pulled out of individual agent runtimes and consolidated into a separate, vendor-neutral tier.
Why it matters
Gartner estimates the average global Fortune 500 company will have more than 150,000 AI agents in use by 2028, up from fewer than 15 in 20251. Yet only 13% of organizations believe they currently have the right AI agent governance in place. ◆ That gap — explosive agent proliferation against near-absent governance — is the demand signal driving a new infrastructure category. The companies and foundations moving fastest are not building better agents; they are building the layer that sits above them.
The big picture
Three developments this month illustrate the separation underway.
First, the standards layer is consolidating. Google's Agent2Agent protocol has officially joined the Agentic AI Foundation (AAIF), the Linux Foundation body that already hosts Anthropic-led Model Context Protocol4,5. A2A enables agents to discover and delegate tasks via "agent cards," while MCP standardizes how AI applications connect to external tools and data. AAIF has grown to more than 250 member companies from fewer than 40 when it launched in December last year, with participants including Google, Microsoft, Amazon, Anthropic, OpenAI, Bloomberg, Shopify, and Block. AAIF Executive Director Marjin Gilbert stressed that moving A2A to AAIF would "provide a neutral space to develop open communication methods between AI agents".
Second, startups are racing to own the runtime governance tier. xpander.ai, founded by three former AWS principal engineers, is making its enterprise AI agent platform generally available, positioning it as a vendor-neutral control plane for building, running, and governing agents across different models, frameworks, and infrastructure environments. CEO David Twizer told VentureBeat that the company increasingly hears three problems from enterprise customers: "agents running locally without centralized governance, agent workflows remaining isolated to individual users, and infrastructure becoming tied to a single AI provider". "The third issue is the most critical part: it's being locked into one vendor," Twizer said. The company announced a $7.5 million seed round led by Pico Venture Partners, with participation from Emerge Ventures, Samsung Next, and SeedIL.
Third, large enterprises are building their own versions of this layer internally. At Capital One, MVP of machine learning engineering Kel Vanee described a centralized, enterprise-wide AI platform with built-in governance, deeply customized open-weight models, and a proprietary multi-agent orchestration harness2. "At Capital One, we're not just using AI, we're building AI," Vanee said. The bank fine-tunes open-weight models with proprietary data rather than relying solely on frontier models. "We view our data as a huge advantage and something that nobody else has, something that the general frontier models cannot provide," Vanee explained.
Between the lines
A VentureBeat Pulse survey of 107 enterprises found that 85% run two or more orchestration platforms3. Microsoft AI Foundry / Copilot Studio appears in 70% of stacks and OpenAI's Agents SDK in 68%, with Anthropic's Claude Platform in 47%. ANALYSIS No single vendor dominates, which makes the case for a vendor-neutral governance layer almost self-evident.
But governance is only half the problem. One in five enterprises still has no real-time way to stop a runaway agent before the bill arrives. ◆ Cost metering — not just permissions or observability — is the gap that current orchestration stacks have not closed. xpander's pricing model, built around individual agent wakes and tool calls rather than seats, is a direct response to this metering vacuum.
Meanwhile, the security implications of inter-agent communication remain unresolved. Mahesh Shanmugasundaram, Seekr's lead AI solutions architect, warned that A2A brings the risk that "unverified claims will travel and gain apparent authority through an agent chain"8. ◆ Placing A2A and MCP under one foundation creates a coordination opportunity for identity and trust standards, but the AAIF transfer itself does not standardize enterprise authorization policies.
Context platforms — governed layers holding metadata, business knowledge, and operating procedures — represent another dimension of the control plane, particularly in industrial settings where the difference is "between a demo that looks useful and an agent that can be trusted" in production6.
ANALYSIS The control plane is becoming the strategic chokepoint. Enterprises selecting orchestration platforms already optimize for flexibility across models rather than affinity to any single provider. As AAIF's membership continues to expand and startups like xpander compete with hyperscaler-native tooling, the question shifts from whether the control plane separates to who owns it — the enterprise, a neutral foundation, or the cloud provider whose infrastructure runs the agents. The next twelve months of standards work at AAIF, and the enterprise procurement decisions that follow, will determine whether agent governance becomes an open layer or a new vector for platform lock-in.