Cisco surveyed its major enterprise customers at RSA Conference 2026 and found something worth sitting with: 85 percent are experimenting with AI agents, but only 5 percent have moved agentic technology into production. That 80-point gap is not a curiosity. It is the central structural problem in enterprise AI right now, and it points to something specific: an orchestration gap between infrastructure layers that each work well in isolation but fail to connect.
The individual layers are real. They are shipping. Forkast has tracked each one as it reached production. The Harness Pattern showed how infrastructure providers bundle orchestration with compute. The shift toward Agent Infrastructure as a Commodity SKU revealed that the competitive moat is moving upstream from raw compute to integrated platform layers. Kubernetes Agent Sandbox became the execution substrate — the unit of agent deployment itself. Execution-Layer Gateways established where enterprise security actually lives in the stack. Okta’s XAA protocol brought agent identity to the provider layer, and the IETF entered the protocol layer to formalize what the market already shipped. Even within enterprise applications, Oracle Fusion embedded orchestration at the ERP layer.
Seven layers. Seven production-grade solutions. And yet the integration between them is where everything breaks down.
The Model Context Protocol governs how agents connect to tools — vertical, agent-to-tool. It explicitly does not handle agent-to-agent communication. The A2A protocol, now under the Linux Foundation with over 150 organizations in production, addresses that horizontal coordination gap. But neither protocol spans the full stack. MCP cannot see what A2A routes. K8s Sandboxes isolate execution but do not enforce identity policy. Execution-Layer Gateways enforce security perimeters but do not coordinate with sandbox lifecycle management. Okta XAA manages identity at the provider layer while the IETF is still formalizing identity at the protocol layer — two overlapping efforts at different stages of maturity with no shared coordination fabric.
Enterprises are left to build custom glue code between these layers, and the cost of that glue is becoming the primary barrier to production deployment. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027 — not because the technology does not work, but because the coordination tax is too high. Anushree Verma, Senior Director Analyst at Gartner, put it directly: “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied. This can blind organizations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production.”
The vendor lock-in fear makes this worse. Between 76 and 81 percent of enterprises express concern about proprietary dependencies in agent memory, model integration, and orchestration tooling. Eighty-seven percent of IT leaders now prioritize interoperability for agentic orchestration. Fifty-one percent prefer hybrid stacks that layer open protocols on top of vendor-managed environments. They want standardized infrastructure without surrendering control to a single provider — but the current stack offers no clean way to achieve that without building the coordination layer themselves.
The number that matters most is this: only 11 to 14 percent of enterprise AI agent pilots reach production at scale. The components are not the problem. Cisco’s DefenseClaw framework, Agent Runtime SDK, and MCP policy enforcement in Secure Access all demonstrate that the individual tools exist. What is missing is the connective tissue — the meta-orchestration layer that enforces policy, manages identity, coordinates execution, and maintains audit trails across the full stack, not just within each layer’s perimeter.
Until someone builds that layer, the 80-point gap between experimentation and production will persist. The infrastructure is ready. The integration is not.