Agent Gateways Are Turning Into Enterprise AI's New Control Layer

Arcade and Manufact have launched agent authorization and MCP hosting products in cloud marketplaces, signaling that agent gateways are emerging as a distinct enterprise category. Nutanix also released its Agent Gateway in May 2026 to provide cost control and governance for autonomous agents, though its MCP server governance remains in tech preview.

Arcade has pushed its agent authorization and tool-execution runtime into the Microsoft Azure and AWS marketplaces, while Manufact is taking MCP hosting from developer workflow to production infrastructure /article/langflow-langgraph-langchain-agent-framework-security-flaws , giving enterprise buyers two fresh signs that agent gateways have become a real category rather than a feature on someone else's roadmap. The shift was laid out in a July 5 Forbes analysis by Janakiram MSV: the software layer between AI agents and the models, APIs, databases, MCP servers and business tools they touch is being carved out into its own control plane. The timing matters. AI agents are moving from chat windows and demos into workflows that spend tokens, open tickets, write to GitHub, query internal systems and invoke third-party services. Once that happens, the real enterprise questions become operational: who sees the traffic, who authorizes the action, who pays for the tokens, and who can stop the agent before it touches the wrong system? Arcade's July 3 marketplace move is the cleanest buyer-side wedge. The company says teams can now deploy its runtime through the Microsoft Azure and AWS marketplaces into their own cloud environments. Arcade's pitch is about giving background agents delegated user authority that is checked when an action is taken. The bet is that enterprise adoption will turn on permissioning, procurement and security review more than on model choice. Manufact is coming from the developer side of the same bottleneck. The Manufact site describes a platform for building and deploying MCP apps and servers; Forbes frames it as taking a Model Context Protocol server from a GitHub push to a monitored production endpoint, with testing across ChatGPT and Claude. Its Y Combinator profile lists the company as a Summer 2025 batch startup. The team is chasing the moment after the demo, when moving to production demands auth, secrets management, monitoring and a stable endpoint. Nutanix gives the category its enterprise shape The incumbent version of the bet comes from Nutanix https://www.nutanix.com/?ref=runtimewire , which shipped Nutanix Agent Gateway https://www.nutanix.com/blog/introducing-nutanix-agent-gateway?ref=runtimewire in late May 2026 as part of Nutanix Enterprise AI 2.7. Nutanix says Agent Gateway is generally available for Enterprise AI customers and is designed to provide cost control and governance for autonomous agents. The product routes traffic from agents to large language models and, through MCP servers, to business tools such as GitHub and Stripe. That is a natural move for Nutanix. The company's current homepage leads with "The Control Layer for Agentic AI at Scale" and puts Agent Gateway at the front of its AI message. The throughline is visible: Nutanix is trying to turn agent sprawl into a platform problem it already knows how to sell to infrastructure teams. The production-readiness caveat is important. Nutanix says Agent Gateway's centralized gateway manages and helps secure traffic from agents to LLMs, and provides usage visibility, policy enforcement and token cost accountability across hosted and self-hosted inference environments. But Nutanix also says its MCP server governance is in Tech Preview, and its own footnote states that Tech Preview features should not be used in production. The generally available core is the routing, observability and rate-limiting layer; the tool-governance piece that makes the agent-gateway story most compelling is still maturing. That split captures the state of the market. Enterprises want a single enforcement point for agents, and vendors are racing to define what should live there. Model routing, token metering, audit logs, tool permissions, MCP federation, identity, runtime protection and policy all fit under the gateway banner. No buyer should assume those capabilities are equally complete across products simply because the category name is settling. The fight is spreading across security, cloud and open source Palo Alto Networks made the consolidation argument explicit in May 2026, when it announced that it had completed the acquisition of Portkey https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-completes-acquisition-of-portkey-to-secure-ai-agents?ref=runtimewire to secure AI agents. The company framed the gateway as a control point for monitoring and governing autonomous agents at scale. Solo.io is pushing the open-source path. In June 2026, the Agentic AI Foundation said agentgateway https://aaif.io/blog/agentgateway-joins-aaif-as-an-open-gateway-for-agentic-ai-infrastructure/?ref=runtimewire had become its fourth hosted project under Linux Foundation governance. The project supports MCP, agent-to-agent traffic, LLM inference, HTTP and gRPC through one layer, and the foundation says it has more than 300 contributors across more than 60 organizations, including CoreWeave, Red Hat, Adobe, Salesforce, Amdocs and Microsoft. Security teams are arriving from the opposite direction: exposure. In January 2026, CyCognito https://www.cycognito.com/blog/introducing-discovery-of-externally-reachable-mcp-services/?ref=runtimewire introduced discovery of externally reachable MCP servers, arguing that exposed endpoints expand attack surface and can reveal sensitive operations. If an agent gateway becomes the place where tool access is declared and enforced, security vendors have every incentive to own that layer or at least instrument it. AWS is already bundling similar concerns into its own stack. The Amazon Bedrock AgentCore FAQ https://aws.amazon.com/bedrock/agentcore/faqs/?ref=runtimewire describes AgentCore Runtime as a secure serverless runtime for agents using open-source frameworks, MCP, agent-to-agent patterns and models inside or outside Bedrock. AWS also describes Agent Registry as a place to discover, reuse and govern agents, tools and agent skills across an organization. For buyers already committed to AWS, that creates a default path. For startups, it raises the bar: a standalone gateway has to prove it offers sharper governance, better cross-cloud neutrality or a faster path to production than the hyperscaler bundle. Why founders are crowding into the gateway The attraction is simple. Agents create a new chokepoint in enterprise software. Every useful agent needs to call models and tools. Every regulated company needs to know which agent called which tool, under whose authority, with what data, at what cost, and with what result. The company that controls that hop can become part of every agent deployment, regardless of which model wins. That is why the teams entering the market are not all building the same thing. Arcade is treating authorization as the wedge. Manufact is treating MCP deployment and operations as the wedge. Nutanix is treating hybrid inference governance as the wedge. Palo Alto is folding gateways into security. Solo.io is turning the same pattern into open infrastructure. AWS is packaging it with runtime, identity and registry services. The category will likely stay messy for a while because the buyer is still being defined. A platform engineering leader may want model routing and observability. A security leader may want runtime protection and attack-surface inventory. A developer tools buyer may want one-click MCP deployment and debugging. A finance team may only care that token burn can be traced to the agent and team that caused it. The winning products will have to cover more than one of those constituencies. A gateway that only routes model calls becomes a cost-management feature. A gateway that only hosts MCP servers becomes deployment tooling. A gateway that only scans exposure becomes a security add-on. The prize is the operational layer that developers can build through and enterprises can govern through. That is why this week's Arcade and Manufact moves matter even beside Nutanix, Palo Alto, AWS and the Linux Foundation. Startups are moving quickly because the control point is still open. Incumbents are moving quickly because they know how often the control point becomes the budget line.