Every enterprise AI architecture conversation in the past six months ends up in the same place: “So who handles the orchestration layer?” The question is rhetorical. Nobody handles it well yet. This is the most important unsolved problem in enterprise AI deployment, and it’s worth being specific about why.
The problem stated plainly
Enterprises are deploying multiple AI agents from multiple vendors: one from Salesforce, one from ServiceNow, one from a specialized vertical vendor, possibly one they built themselves. Each agent is optimized for a specific task. None of them know what the others are doing. And the enterprise needs them to coordinate — passing context, handing off tasks, reconciling outputs — in real time, across data boundaries, without losing audit trail.
This is the agent-of-agents problem. It’s not a new problem conceptually. It’s a very hard problem practically, because it requires standards, trust models, and latency characteristics that don’t exist yet at enterprise scale.
Why MCP isn’t enough
Anthropic’s Model Context Protocol (MCP) is a good tool for a different job. MCP is about how a single agent connects to tools and data sources. It’s a plug standard for agent-to-tool communication. It does not address agent-to-agent communication, which is the harder problem.
The confusion in the market is real. Several vendors are pitching “MCP support” as the answer to enterprise orchestration. It isn’t. MCP helps an agent call your CRM API. It doesn’t help two agents from different vendors decide who owns a customer record or which agent’s conclusion wins when they produce conflicting outputs.
Why A2A might be closer
Google’s Agent-to-Agent (A2A) protocol, which has been circulating in draft form, takes the right approach at the protocol level. It defines how agents advertise their capabilities, how they negotiate task handoffs, and how they establish trust without requiring a single central orchestrator. The draft is promising. The adoption is nascent.
The challenge is that A2A requires buy-in from multiple vendors to be useful. That’s a standards problem as much as a technical problem, and standards problems in enterprise software typically take 3–5 years to resolve. A2A will get there. It won’t be the answer in 2026.
What enterprises are doing in the meantime
The pragmatic answer, and the one we see most often in the deployments we watch closely, is a thin proprietary orchestration layer: a lightweight process that routes work to the right agent, maintains shared context in a simple key-value store, and logs everything. It’s not elegant. It’s also not MCP or A2A. It’s a custom integration with a clear human ownership model and a budget for maintaining it.
The company that builds the clean, standardized, vendor-neutral version of that layer — and gets two or three major platform vendors to endorse it — wins a very large market. Nobody has done it yet.