AWS, Azure, and GCP have each built sophisticated orchestration machinery within their own perimeters. The gaps that remain aren't oversights; they're the direct consequence of building agent infrastructure optimized for platform retention. AI agent orchestration across the cloud surfaces three structural problems that no single provider resolves.
Agent Discovery Stops at the Cloud Boundary
Every major platform publishes agents within its own registry. Bedrock AgentCore exposes agents through AWS service endpoints. Foundry Agent Service registers agents within the Azure tenant. Vertex AI Agent Engine surfaces agents through GCP project namespaces. None of these registries federates across cloud boundaries.
A2A's Agent Card specification was designed to address exactly this by allowing agents to advertise their capabilities at a well-known URL. In practice, a Bedrock supervisor agent has no native mechanism to query a GCP-hosted Agent Card, evaluate its trustworthiness, or route tasks to it without custom engineering. Discovery requires more than a specification; it requires a live, provider-neutral registry that all three clouds can read and write to.
For enterprises running AI agent orchestration across cloud environments, the practical consequence is that cross-cloud agent graphs get encoded as hardcoded endpoint tables maintained by engineering teams. When agents change, those tables drift. When agents fail, nothing automatically reroutes.
Delegation Chains Break Across Trust Domains
The delegation problem runs deeper than credentials. When a Bedrock supervisor delegates a subtask to an Azure-hosted agent, the delegation event carries no portable authority token. AWS IAM, Entra ID, and GCP's Workload Identity Federation operate as separate trust anchors with no shared attestation model to verify that the delegating agent had authority to delegate, that the receiving agent is the intended recipient, or that the task scope hasn't been altered in transit.
MCP adopted OAuth 2.1 as an optional authorization layer, yet security scans of MCP servers found that a large minority lacked authentication entirely. A2A agent cards contain self-declared identities with no attestation binding. Cloud-agnostic AI agent orchestration solutions require verifiable delegation, not just routable task envelopes.
No Neutral Coordination Plane Exists
The deepest gap is architectural. Bedrock's multi-agent collaboration routes are implemented via AWS Lambda and Step Functions. Foundry's orchestration runs through Azure Durable Functions and the Microsoft Agent Framework. GCP's coordination relies on Vertex AI Agent Engine and Apigee. Each is coherent within its domain. None provides a neutral plane where agents across all three clouds participate in the same governed workflow with shared state, unified observability, and consistent policy enforcement.
Best agent orchestration for hybrid cloud RAG workloads makes this concrete. When retrieval pipelines span Bedrock Knowledge Bases, Foundry IQ, and Vertex AI RAG Engine simultaneously, there's no cross-cloud ranking layer, no shared relevance model, and no governance checkpoint controlling which retrieved context an agent is authorized to reason over.
Proprietary extensions to A2A and MCP accelerate the divergence. Every platform-specific addition is another surface where cross-cloud coordination breaks down without a neutral intermediary to normalize it.