MCP’s meteoric rise spurs push for unified agent protocols

The October 2026 MCP Dev Summit’s report of 500 million monthly SDK downloads landed as evidence that adoption had outrun the industry’s original plumbing. Prior to November 25, 2024, AI applications connected models to external systems through application-specific tool-calling code, involving custom API wrappers, adapters, and authentication handling unique to each application, so the summit’s scale milestone underscored not just success for MCP but how badly developers still needed a broader coordination layer once standardized tool access became mainstream.

That urgency was even clearer because the fragmentation problem had never been limited to back-end tool connectivity alone. Before May 12, 2025, agent-backed user interfaces relied on custom WebSocket protocols, text parsing, and framework-specific frontend integrations to communicate between agents and applications, meaning the summit’s headline growth also highlighted a remaining gap: MCP had proven demand for common connections, but production systems still needed a way to coordinate across multiple protocol domains rather than standardize only one layer.

Standardization is accelerating because MCP solved the integration problem in a way enterprises and developers could reuse. Designing with AI describes it as an open standard for connecting AI assistants to external data systems, where teams can write a tool once as an MCP server and any client that supports the protocol can use it. Startup Fortune called MCP the open standard Anthropic released in November 2024 that lets AI models plug into tools and data without custom integrations for each one.

That design became a true ecosystem standard as adoption and governance converged. Startup Fortune reported that OpenAI added full support across its Agents SDK, Responses API, and ChatGPT desktop app in March 2025, and by December 2025 Anthropic had handed governance to the Linux Foundation, with OpenAI, Google, and Microsoft as co-sponsors. Nerq counted 17,468 public MCP servers by early 2026, while one analysis noted 15 competing standards and said this one had become the standard. Anthropic’s July overhaul added a formal 12 month deprecation policy and tighter security, reinforcing the shift toward governed scale.

The emerging agent stack is no longer synonymous with MCP because MCP standardizes how agents reach tools and data, while other protocols now cover the layers MCP does not. Daily Dose of Data Science describes the split directly: MCP is “how agents connect to tools, data, and workflows,” AG-UI is “the bi-directional connection between agentic backends and frontends,” and A2A is “the protocol for multi-agent coordination,” a distinction reinforced by the rise of function calling after mid-2023 and the April 2024 shift away from plugins toward explicitly defined tool access. That widening role for MCP also sits alongside its scale, with over 10,000 published servers at the time.

What completes the stack is that AG-UI turns agent outputs into interactive application surfaces, while A2A lets separate agents discover one another, communicate, and hand off work across systems. Syntax notes MCP by itself generally returns data “in text,” whereas AG-UI-style systems return renderable interface elements; AWS similarly defines AG-UI as standardized streaming for agent-to-user communication, and Ground Level AI notes Google “followed in April 2025 with Agent2Agent, or A2A,” enabling cases where “a company’s procurement agent” can coordinate with a supplier’s agent—evidence that the operative stack now spans tool access, user interface, and agent orchestration.

The enterprise bottleneck in agent deployments is no longer whether models can reason, but whether systems can coordinate actions under control. Forkast News cites the UC Berkeley MAST taxonomy finding that 79% of multi-agent failures trace to specification problems rather than model limitations, while Gartner warns 40% of autonomous AI efforts will be partially derailed by governance gaps discovered only after production incidents; in practice, agents fail when authority, delegation, and policy are ambiguous, not when the model lacks fluency. That trust deficit is visible beyond large enterprises: the National AI Centre's latest SME AI Pulse found around 65% of Australian businesses not adopting AI cited either distrust in AI decision-making or a strong preference to keep humans in control.

That is why enterprises are moving from exposed endpoints to governed capabilities: on August 19, Salesforce expanded Headless 360 to turn its clouds into reusable enterprise capabilities that authorized AI agents can securely discover and use through open standards. Salesforce said most MCP servers merely expose APIs, while its Headless 360 MCP Server exposes trusted, metadata-aware business capabilities so agents understand relationships, permissions, and workflows; with more than 100 reusable Agent Skills framed as production software assets, not prompt snippets, the message is that identity, audit trails, and governance now determine scale. The urgency is reinforced by reports of 87% of organizations reporting an API security incident in 2025, with enforcement failures linked to the kind of escalation seen in the July 2026 Hugging Face incident.

What is changing is not just protocol adoption but who is choosing to standardize together. As TechTarget reported, Salesforce’s Gary Lerhaupt said vendors were effectively saying, “Let’s all standardize on these things and not create 15 variants of the same type of job to be done,” while A2A “is supported by over 150 organizations” and has reached version 1.0 with integrations across Google, Microsoft, and AWS; AI Magazine adds that A2A has “more than 150 partner organisations” and sits under Linux Foundation governance, with AWS, Cisco, Microsoft, Salesforce and ServiceNow among AAIF’s founding members.

That alignment is reinforced by institutions and open source choosing consolidation over parallel stacks. AI Magazine noted that “In August 2025, IBM's Agent Communication Protocol merged into A2A,” while Forkast said AAIF “serves as the hub for interoperability and open-source infrastructure,” managing MCP, goose, and AGENTS.md and growing from 146 members in February 2026 to over 290 by August; at the security layer, Resilient Cyber reported that “Dick Hardt, who authored the OAuth 2.0 RFC, designed AAuth specifically because the architecture has changed,” and Forkast added that OWASP ACS “focuses on the runtime enforcement layer” with vendor-agnostic middleware hooks.