Part 6 of the series "The Agent Coordination Gap"
Why Hyperscalers Have No Incentive to Solve Cross-Vendor Agent Coordination, and What Enterprises Should Do About It
The Interoperability That Wasn't
A global insurance company deployed AI agents across four platforms in early 2026. Customer-facing agents ran on Salesforce Agentforce. Claims processing agents operated on ServiceNow. Financial reconciliation agents ran on SAP through Joule Studio. And a cross-functional orchestration layer sat on Microsoft's Copilot Studio, chosen specifically because Microsoft marketed it as the coordination hub that could tie everything together.
On paper, the architecture was sound. Each platform handled what it did best. The Microsoft orchestration layer would coordinate across boundaries, applying governance policies and routing work between agents on different platforms.
In practice, the coordination broke down within weeks. The Salesforce agents exposed their capabilities through MCP, but the Microsoft orchestration layer's governance policies did not extend to actions those agents took inside the Salesforce platform. The SAP agents communicated with Copilot Studio through A2A protocol, but the bidirectional agent identity model had not reached general availability, so the orchestration layer could not verify the SAP agent's permissions against the company's enterprise identity framework. The ServiceNow AI Control Tower, meanwhile, offered its own cross-platform governance, creating a competing coordination layer that overlapped with and sometimes contradicted the Microsoft one.
The company had not built a multi-agent system. It had built four single-vendor agent systems connected by protocols that handled message passing but not coordination. The governance gap between platforms was not a technical limitation waiting for a patch. It was a structural feature of how each vendor designed its agent platform.
This is the platform lock-in trap. Every major vendor is building an agent platform designed to be the coordination layer for the enterprise's agent estate. Each platform works well within its own boundaries. None has a commercial incentive to make cross-vendor coordination work as well as within-vendor coordination. And the enterprise that picks one vendor's platform as its coordination layer discovers that it has traded the integration problem for a lock-in problem that is harder and more expensive to escape.
The Vendor Strategies: Openness as a Moat
Each major vendor has adopted a version of the same strategy: appear open at the model and protocol layer while locking in the platform, orchestration, and governance layers where the durable commercial value sits. The specifics differ, but the pattern is consistent.
Microsoft has executed this strategy most explicitly. Copilot Studio's August 2026 rebuild brought A2A protocol support to general availability, MCP-enabled tools into preview, and third-party model support including Claude. The message is openness: bring any model, connect any agent, use any protocol. But the orchestration layer, the Workflow Designer, runs on Microsoft's platform. Agent identities flow through Entra. Governance policies are defined in Microsoft's security fabric. The models are interchangeable. The platform is not. Microsoft's strategic shift from model lock-in to platform lock-in is deliberate and well-executed. An enterprise that builds its agent coordination on Copilot Studio can swap out GPT-4o for Claude without disruption. Swapping out the Microsoft orchestration layer itself would require rebuilding every workflow, every governance policy, and every identity integration from scratch.
Salesforce has taken a similar path with a different emphasis. Agentforce now includes native MCP support. The Headless 360 initiative exposes more than 60 MCP tools, positioning Salesforce data and processes as accessible to any MCP-compatible agent. The AgentExchange marketplace invites third-party agent builders onto the platform. Agentforce 360 positions Salesforce as the agent operating system for the enterprise. The openness is real at the connectivity layer. The lock-in is equally real at the platform layer: customer data, workflow logic, agent configurations, and the marketplace ecosystem all create switching costs that compound over time. Industry analysts have flagged this pattern, warning that despite open ecosystem claims, the proprietary platform creates dependencies that are difficult to unwind.
AWS has pursued what looks like the most open approach. Bedrock AgentCore supports a "bring-your-own-framework" model: LangGraph, CrewAI, Google ADK, and other open-source frameworks run on the platform. The July 2026 decision to stop accepting new customers on Bedrock Agents Classic in favor of AgentCore signaled a commitment to framework portability. But the platform layer is still AWS. The compute, storage, monitoring, and deployment infrastructure are all AWS services. Framework portability within a single cloud provider is not the same as platform portability across cloud providers. An enterprise running LangGraph on AgentCore can, in theory, move that LangGraph code to another infrastructure. In practice, the integrations with AWS services, the monitoring dashboards, the deployment pipelines, and the data connections all have to be rebuilt. Bedrock Managed Agents, the more opinionated counterpart, makes the lock-in explicit by optimizing for simplicity within the AWS ecosystem.
ServiceNow has positioned itself differently from the other three. The AI Control Tower is marketed as a governance layer for agents across the enterprise, regardless of which cloud or model they run on. The "Super Intelligence Control Tower" framing extends governance to agents on AWS, Google Cloud, and Azure. The Control Tower Alliance brings in partners. Acquisitions of Veza for identity governance and Armis for asset intelligence add capabilities that work across platforms. ServiceNow's deeper integration with Microsoft Agent 365 suggests the company sees its role as the governance overlay, not the execution platform.
ServiceNow comes closest to a vendor-neutral coordination layer. But "closest" is not "neutral." The governance policies, the agent registry, the audit infrastructure, and the control tower interface all run on ServiceNow. An enterprise that builds its agent governance on ServiceNow's AI Control Tower has made ServiceNow the indispensable layer in its agent architecture. The $3.67 billion in quarterly subscription revenue and 22 percent year-over-year growth suggest the market is responding to this positioning. Whether it delivers genuine cross-platform governance or creates a new form of platform dependency is the open question.
The SAP-Microsoft Test Case
The SAP Joule Studio and Microsoft Copilot Studio integration is the most visible test of whether two major vendors can build genuine cross-platform agent coordination. It deserves close examination because it shows both the promise and the structural limits of bilateral vendor integration.
The integration offers two paths. A managed option runs SAP agents through Copilot Studio directly. The alternative routes through an Agent Gateway that uses A2A protocol for bidirectional communication between Joule and Copilot agents. The A2A path is the more interesting one architecturally, because it allows agents on each platform to discover and invoke capabilities on the other without either platform subordinating itself.
The governance challenges are where the integration gets complicated. When a Copilot agent invokes a Joule agent to process a purchase order, whose identity framework governs the transaction? Microsoft's Entra or SAP's own identity layer? When the Joule agent takes an action that violates a governance policy defined in the Microsoft orchestration layer, which platform's audit trail captures the violation? The Agent Gateway that would bridge these concerns has not yet been confirmed as generally available.
This is not a criticism of either vendor's engineering. The SAP-Microsoft integration is more ambitious than anything else in the market. The point is structural: even when two vendors with aligned commercial interests (SAP and Microsoft have a deep strategic partnership) invest in building cross-platform agent coordination, the governance, identity, and audit challenges are substantial. And bilateral integration between two willing partners does not scale to an enterprise running agents on four or five platforms. Each new vendor pair introduces another bilateral integration to build and maintain.
The arbitration problem described in Part 4 of this series applies directly. When a SAP agent and a Microsoft agent take conflicting actions, the arbitration mechanism must operate across both platforms. Neither platform's internal conflict resolution works across the boundary. The bilateral integration provides connectivity. It does not provide arbitration, and it does not provide unified governance.
The Economics of Lock-In
The financial data on platform lock-in makes the strategic case for enterprise-owned coordination layers concrete.
CloudBees' 2025 research found that 57% of cloud migrations exceed $1 million in cost. Arion Research analysis of enterprise agent deployments puts switching costs above $315,000 per project, with that figure rising as agent deployments mature and accumulate platform-specific dependencies. Only 6% of organizations report the ability to change AI vendors without major disruption. Industry estimates place the total switching cost at three to five times the annual vendor spend for the platform being replaced.
The concern is widespread and growing. Between 76 and 81% of enterprises report concern about AI platform lock-in, depending on the survey. The concern is well-founded, because lock-in in agent platforms concentrates in three areas that are harder to migrate than traditional software dependencies.
Agent memory and learned behavior. An agent that has been operating for months accumulates context: interaction histories, learned preferences, calibrated decision parameters, and performance data that improves its effectiveness over time. This accumulated intelligence is stored in the vendor's platform in proprietary formats. Moving the agent to a new platform means losing the operational learning, resetting to baseline, and rebuilding from scratch. The learning flywheel described in the AI Strategy series depends on continuous accumulation. Platform migration breaks it.
Model integration and prompt engineering. Each vendor's agent platform optimizes for specific model families and interaction patterns. Prompt templates, guardrails, output parsers, and evaluation frameworks are all tuned to the platform's preferred models. Moving to a different platform often requires re-engineering these components, even when the underlying model is the same, because the platform's orchestration layer mediates every model interaction.
Orchestration and governance tooling. This is the deepest lock-in layer. The workflow definitions, governance policies, decision authority rules, escalation paths, and audit integrations that the Operating Model Gap series identified as essential to AI value creation all live in the vendor's platform. These are not code artifacts that can be exported and re-imported. They are configurations embedded in a platform-specific framework, referencing platform-specific identity systems, platform-specific data connectors, and platform-specific monitoring infrastructure. Rebuilding them on a new platform is not migration. It is reimplementation.
The total cost of ownership implications compound the switching cost problem. Maintenance and compliance costs run between 20 and 50% of total cost of ownership for agent deployments. These costs are lower when the agent estate stays within a single vendor's platform, because the vendor handles infrastructure updates, security patches, and compliance certifications. Moving to a multi-vendor architecture with an enterprise-owned coordination layer increases these operational costs. The question is whether the increased operational cost is justified by the strategic flexibility it provides. For enterprises running agents on multiple platforms, the answer is almost always yes, because the alternative is a coordination layer controlled by one vendor that the other vendors have no incentive to support.
The Open Alternative: Protocols, Gateways, and What Is Missing
The open-source and standards community has responded to the platform lock-in problem with protocols, frameworks, and gateway projects. The question is whether these alternatives are mature enough to serve as the foundation for an enterprise-owned coordination layer.
The protocol landscape is converging. MCP handles tool and data access: an agent uses MCP to invoke a tool or query a data source. A2A handles agent-to-agent communication: agents use A2A to discover each other's capabilities and coordinate on tasks. The two protocols are complementary, and both are now under the Linux Foundation's AI and AI Foundation (AAIF), which has grown to 146 member organizations. MCP is MIT-licensed. A2A is Apache 2.0. Neither protocol is controlled by a single vendor, though Google originated A2A and Anthropic originated MCP.
Beyond MCP and A2A, the Agent Communication Protocol (ACP) from IBM and the BeeAI project takes a REST-first approach to agent interoperability. AGNTCY, backed by Cisco and hosted at the Linux Foundation, provides an infrastructure-layer framework for agent networking. The protocol options are proliferating, which is both a sign of market demand and a source of fragmentation risk.
Open-source gateway projects are emerging to bridge the gap between protocols and platforms. Bifrost provides a universal gateway for agent communication across protocols. The agentgateway project offers similar bridging capabilities. IBM's ContextForge focuses on context management across agent boundaries. Envoy AI Gateway extends the established Envoy proxy project to handle AI agent traffic. These projects are early-stage, but they point toward an architecture where the coordination layer is a piece of infrastructure the enterprise operates, not a platform a vendor controls.
What the open alternatives lack is the enterprise governance layer that the five functions of the Agent Service Bus framework require. Protocols handle connectivity. Gateways handle routing. Neither provides the enterprise agent registry, coordination-layer policy enforcement, cross-platform audit integration, proportional governance, or governance board authority that Part 5 of this series identified as requirements for multi-agent governance at scale. The governance gap between what open protocols deliver and what enterprises need is the most significant barrier to enterprise-owned coordination layers today.
This gap is where the organizational dimension becomes decisive. The Operating Model Gap series established that 70% of AI value comes from people, process, and behavior change. The coordination layer is not just technology. It is an organizational capability that includes decision authority frameworks, hub-and-spoke governance structures, and proportional governance policies that match governance intensity to risk. An enterprise building its own coordination layer must build the organizational infrastructure alongside the technical one. Only 3% of organizations are successfully scaling multi-agent systems across departments, and only about 6% are capturing meaningful value from AI at scale. The coordination layer challenge is a primary reason for both numbers.
Building the Enterprise-Owned Coordination Layer
The case for an enterprise-owned coordination layer does not require building everything from scratch. It requires a deliberate architecture that uses vendor platforms for execution while keeping coordination, governance, and arbitration under enterprise control.
The architectural pattern maps to the Agent Service Bus framework from earlier articles in this series. The five ASB functions (identity and trust, message routing, state management, conflict arbitration, and policy enforcement) form the coordination layer. Vendor platforms handle agent execution within their domains. Open protocols (MCP, A2A) handle communication between domains. The enterprise-owned layer handles everything that sits above communication: deciding which agent has authority, applying governance policies, detecting and resolving conflicts, and maintaining the audit trail.
The three-tier decision authority model from the Operating Model Gap series applies at the coordination layer level. Tier 1 coordination (routine, pre-authorized cross-platform interactions) can be automated through the gateway and policy engine. Tier 2 coordination (interactions that require review or involve elevated risk) routes to human decision-makers with full context from both platforms. Tier 3 coordination (strategic decisions about which vendor platforms to use, how to allocate workloads, and when to migrate) stays with enterprise leadership.
The Dual Maturity Framework provides the assessment model. Organizations at Level 1 (Ad Hoc) maturity should not attempt an enterprise-owned coordination layer. They lack the governance structures, the organizational capabilities, and the technical infrastructure to operate one. Organizations at Level 2 (Managed) can begin with bilateral coordination between their two most critical platforms, using open protocols and a lightweight gateway. Organizations at Level 3 (Systemic) can build the full enterprise-owned coordination layer, with the ASB framework as the architectural guide and the hub-and-spoke governance model as the organizational foundation.
The human-in-the-lead principle from throughout both the Operating Model Gap and Agent Coordination Gap series is especially important here. The coordination layer is where enterprise control over the agent estate is exercised. If that layer is owned by a vendor, the enterprise's control is mediated by the vendor's commercial interests. If that layer is owned by the enterprise, human leaders retain direct authority over how agents coordinate, what governance policies apply, and how conflicts between vendor platforms are resolved.
What Comes Next
The platform lock-in trap is a structural problem within a single enterprise. The next article in this series examines what happens when multi-agent coordination crosses organizational boundaries: the inter-enterprise coordination problem. When agents from different companies must interact, negotiate, and transact, there is no shared governance framework, no common trust boundary, and no single coordination layer that spans both organizations. The inter-enterprise problem is harder than the intra-enterprise problem by an order of magnitude, and no vendor, open-source project, or standards body has a credible answer yet.
Strategy Playbook
1. Platform Dependency Audit. Map every agent in your estate to the vendor platform it runs on. For each platform, inventory three categories of dependency: agent memory and learned behavior (what would be lost in a migration), model integration assets (prompt templates, guardrails, evaluation frameworks), and orchestration configurations (workflow definitions, governance policies, identity integrations). Score each category on a 1 to 4 scale of migration difficulty. The platforms with the highest aggregate scores are your deepest lock-in points. Conduct this audit quarterly, because lock-in deepens as agents accumulate operational data and as teams build more workflows on each platform.
2. Coordination Layer Ownership Assessment. Determine which vendor currently owns your de facto coordination layer, the system that decides how agents on different platforms interact. If a single vendor controls this layer, assess the business risk: what happens if that vendor changes pricing, deprecates a feature, or is acquired? Define the target state: which coordination functions should the enterprise own, and which can remain with vendors? The five ASB functions from earlier in this series provide the framework. At minimum, policy enforcement and conflict arbitration should be under enterprise control, even if identity, routing, and state management leverage vendor infrastructure.
3. Open Protocol Adoption Roadmap. Evaluate MCP and A2A adoption across your vendor platforms. Where vendors support these protocols natively (Salesforce's MCP support, Microsoft's A2A support), use them as the communication layer between platforms rather than vendor-specific integration APIs. Where vendors do not yet support open protocols, assess open-source gateway options (Bifrost, agentgateway, Envoy AI Gateway) as bridging mechanisms. Build internal expertise on the protocols themselves, not just on vendor implementations of them. Protocol expertise is portable across vendors. Vendor expertise is not.
4. Bilateral Coordination Pilot. Select the two vendor platforms in your estate that have the most critical cross-platform workflows. Build a coordination proof of concept using open protocols for communication and a lightweight enterprise-owned policy layer for governance. The SAP-Microsoft A2A integration pattern is a reference architecture, not because it is complete, but because it illustrates the right structural approach: agents communicate through open protocols while governance remains under organizational control. Measure the pilot against three criteria: does the coordination layer enforce governance policies across both platforms, can it detect and escalate conflicts between agents on different platforms, and does the audit trail capture cross-platform interactions with enough detail for compliance review? Use the pilot results to build the business case for broader coordination layer investment.
Arion Research advises enterprise leaders on AI strategy and the shift to a digital workforce.