By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 2, 2026, 16:19
Salesforce just reported $800 million in Agentforce ARR — up 169% year-over-year — with 18,500 customers and 2.4 billion Agentic Work Units delivered. Behind those numbers is an architectural decision that matters more than the revenue: Salesforce adopted two open protocols, MCP and A2A, as the backbone for how its AI agents connect to the world. This is a company that has historically preferred to own every layer of its stack. The fact that they are building on open standards tells you something about where enterprise AI is heading.
Two Protocols, Two Problems
To understand what Salesforce is doing, you need to understand the two protocols and the distinct problems they solve.
Model Context Protocol (MCP) is the agent-to-tool connection. Originally developed by Anthropic, MCP defines how an AI agent discovers and uses external tools — databases, APIs, file systems, SaaS products. Think of it as USB-C for AI: one standard plug that connects any agent to any tool. Before MCP, every integration was custom. Every AI vendor built its own connectors to every data source. MCP replaces that N-times-M integration problem with a single standard.
Agent-to-Agent Protocol (A2A) is the agent-to-agent connection. Launched by Google in April 2025 with over 50 technology partners (Salesforce among them), A2A defines how AI agents discover each other, exchange tasks, and collaborate. Each agent publishes a card describing its capabilities. Other agents read that card, send tasks, and receive results — all over standard JSON-RPC 2.0 and HTTP.
MCP lets an agent use tools. A2A lets agents use each other. Together, they create a world where AI agents are not isolated applications but networked participants in a shared ecosystem.
What Salesforce Built
Agentforce 3 and Native MCP
Agentforce 3, launched in June 2025, introduced native MCP client support — entering pilot in July 2025. This means any Agentforce agent can connect to any MCP-compliant server without custom code. The implications are significant:
- A Salesforce agent can reach into external systems — AWS databases, Box file repositories, Stripe payment processing, Google Maps — through standardized MCP connections
- Partners build one MCP server for their tool, and every Agentforce customer gets access
- No vendor lock-in. The same MCP server that works with Agentforce works with Claude, with Cursor, with any MCP-compatible client
Salesforce ships three production-ready MCP servers today:
- Salesforce DX MCP Server — natural language commands for deployments, scratch org creation, and test execution
- Heroku Platform MCP Server — manage apps, dynos, and add-ons through conversational prompts
- MuleSoft MCP Server — MuleSoft project management and deployment via AI
The enterprise layer on top is what distinguishes Salesforce’s implementation from a developer spinning up an MCP server in a weekend. Agentforce provides a centralized MCP server registry where administrators enforce security policies, rate limiting, identity verification, and access controls. Every agent connection goes through governance. This is the “trust layer” that enterprises require and that bare-metal MCP does not provide.
A2A: Agents Talking to Agents
Salesforce joined the Linux Foundation’s A2A Project as a founding member, committing to implement the protocol across Agentforce. The vision: a Salesforce CRM agent delegates a compliance check to a ServiceNow agent, which delegates a document retrieval to a Box agent, which returns the signed NDA with key term summaries — all through standardized A2A task exchanges.
MuleSoft’s Agent Fabric — Salesforce’s integration middleware — added Agent Scanners that automatically extract capabilities from existing agents and map them to A2A agent card specifications. This means organizations can take their current AI agents, whatever framework they are built on, and expose them as A2A-discoverable services without rewriting anything.
The practical state of A2A in production is more cautious than the announcements suggest. A production-ready version of the protocol was targeted for late 2025 but has not fully materialized. Most partnerships remain at the announcement stage rather than verified production deployments. Salesforce is building the plumbing, but the water is not fully flowing yet.
The Enterprise Reality: 12 Agents and Counting
Salesforce’s 2026 Connectivity Report surveyed enterprise IT leaders and found numbers that explain why these protocols matter:
- Organizations currently deploy an average of 12 AI agents, projected to grow 67% by 2027
- 83% of organizations report most or all teams have adopted AI agents
- 50% of those agents operate in isolated silos — not connected to each other or to a unified system
- 96% of IT leaders say agent success depends on seamless data integration across systems
- 27% of APIs remain ungoverned on average
- Only 54% maintain centralized governance for agentic capabilities
The picture is clear: enterprises adopted AI agents fast, but adopted them in fragments. Every department bought or built its own agent. Now they have a dozen disconnected AI systems that cannot share data, cannot delegate tasks to each other, and cannot be governed from a central point.
This is exactly the problem MCP and A2A were designed to solve. MCP standardizes how agents access tools and data. A2A standardizes how agents collaborate. Without these protocols, every pair of agents needs a custom integration. With them, you build once and connect everywhere.
Protocol adoption among enterprises reflects growing interest but not yet universal commitment:
| Protocol | Organizations Supporting or Planning |
|---|---|
| Agent Network Protocol | 43% |
| Agent Communication Protocol | 43% |
| Agent-to-Agent (A2A) | 40% |
| Model Context Protocol (MCP) | 39% |
| Universal Tool Calling Protocol | 34% |
No single protocol has broken away. The market is still deciding. But MCP and A2A, backed by Anthropic and Google respectively, have the strongest momentum and the broadest ecosystem support.
The Partner Ecosystem
Salesforce’s AgentExchange now lists over 30 MCP integration partners. The use cases show the breadth of what becomes possible when tools are standardized:
AWS — Agents access unstructured data repositories, summarize documents, extract information from images, analyze audio and video, and query databases through natural language. All through MCP, no custom AWS SDK integration required.
Box — A Salesforce agent retrieves signed NDAs, generates key term summaries, and posts compliance recommendations to Slack. Three systems, one agent interaction, zero custom code.
PayPal and Stripe — Full commerce capabilities: product listings, orders, payments, subscriptions, refunds, dispute management. A customer service agent that can actually process a refund instead of telling the customer to call another department.
Google Cloud — Agents connect to Maps Platform for location-based services and to generative models (Veo, Imagen, Lyria) for content creation. A marketing agent that generates localized content with accurate mapping data.
Each partner built one MCP server. That server works with every Agentforce customer. It also works with Claude Desktop, with VS Code extensions, with any MCP client. The economics of this are powerful — build once, reach every AI platform that speaks the protocol.
Why This Matters Beyond Salesforce
Salesforce is not the only company adopting these protocols. They are one of the largest, which makes their decisions influential, but the significance is in the pattern, not the brand.
The integration tax is real. Before MCP, connecting an AI agent to a CRM required a custom integration. Connecting it to a payment system required another. To a document store, another. Each integration cost engineering time, introduced maintenance burden, and created a point of failure. MCP collapses that cost. One protocol, one integration pattern, every tool.
Agent silos are the new data silos. The 2026 Connectivity Report finding that 50% of agents operate in isolation is the AI equivalent of the data silo problem that has plagued enterprises for decades. A2A is the proposed solution — the same way APIs standardized how applications talk to each other, A2A standardizes how agents talk to each other.
Governance cannot be an afterthought. With 27% of APIs ungoverned and only 54% of organizations maintaining centralized agentic governance, the security surface area is expanding faster than the controls. Salesforce’s enterprise MCP registry — centralized policy enforcement, rate limiting, access control — is their answer. It is also a competitive moat: open-source MCP is free, but enterprise governance on top of it is a Salesforce product.
The 16x claim. Salesforce states that organizations can deploy autonomous AI agents 16 times faster using Agentforce with MCP than traditional DIY approaches. The number is marketing, but the direction is credible. Standardized protocols eliminate boilerplate. When every integration follows the same pattern, the first one takes time and the next twenty are fast.
The Honest Assessment
Salesforce is making the right architectural bet. Open protocols for agent connectivity, proprietary governance on top. This is the same playbook they used with APIs a decade ago — Salesforce did not invent REST, but they built the most successful enterprise platform on top of it.
The risks are real:
A2A is still early. The protocol launched in April 2025. Production deployments are sparse. The Linux Foundation governance is a good sign for longevity, but the specification is still evolving. Building on A2A today means building on a moving target.
Protocol fragmentation. Five competing protocols, none above 43% adoption. The market may not converge on MCP and A2A. It may converge on something else, or it may fragment into multiple standards — which would defeat the purpose entirely.
The governance tax. Salesforce’s enterprise MCP registry adds security and control, but it also adds cost and complexity. Organizations that do not need enterprise-grade governance — startups, small teams, internal tools — may find the overhead unjustified. The open-source versions of MCP and A2A work fine without a centralized registry.
Vendor dependency. Salesforce is building on open protocols, but the governance layer, the AgentExchange marketplace, the MuleSoft Agent Fabric — these are proprietary. An organization that invests deeply in Salesforce’s implementation of MCP will find it harder, not easier, to leave the Salesforce ecosystem. Open at the protocol layer, locked at the platform layer.
What Comes Next
The next 12 months will determine whether MCP and A2A become the HTTP of AI agents or the SOAP of AI agents — a good idea that gets replaced by something simpler.
The signals are positive. Anthropic continues to invest in MCP. Google and 50+ partners back A2A through the Linux Foundation. Salesforce, the largest CRM company in the world, is building its entire agent platform on both. AWS, Google Cloud, and Microsoft all have MCP support in progress.
But the real test is not vendor announcements. It is production deployments. It is a mid-size company running a Salesforce CRM agent that delegates tasks to a custom Python agent over A2A, using MCP to access a dozen different data sources, governed by a central policy, and doing it reliably at scale. That scenario is architecturally possible today. Whether it is practically achievable — with acceptable latency, failure handling, and operational complexity — remains to be proven.
The enterprise has 12 agents and growing. Those agents need to talk to tools and to each other. MCP and A2A are the best candidates for making that work. Salesforce is betting $800 million in ARR on it. The bet looks sound. The execution is what matters now.