Agentic AI
Big Cloud Providers Converge on Enterprise Agent Blueprints, But Portability Remains in Limbo
Amazon, Microsoft, and Google have independently arrived at the same architectural components for production AI agents—but lack a unified standard to move workloads between clouds.
The New Stack
Within the past year, the three major cloud vendors have each launched or revamped an enterprise agent platform using nearly identical building blocks: runtime environments, memory management, tool gateways, identity systems, observability layers, and governance controls. Yet despite these striking similarities across Amazon Bedrock AgentCore, Microsoft Foundry, and the Gemini Enterprise Agent Platform, operators face a familiar problem: each solution chains them to a single provider's infrastructure.
What unfolded between 2011 and 2016 with cloud infrastructure provides an instructive parallel. Back then, building an application meant stitching together virtual machines, load balancers, message queues, and monitoring tools—each with its own API and operational quirks. Cloud Foundry and Heroku changed that by establishing a common contract between applications and infrastructure. Developers could reason about portability rather than plumbing. The agent ecosystem now stands at a similar threshold, but without the open standard to unlock it.
The Architecture Taking Shape#
The convergence across the three platforms reveals clear patterns. Amazon's AgentCore offers runtime environments with eight-hour execution windows, a tool gateway compatible with MCP servers, session-isolated memory, and observability exports via OpenTelemetry. Microsoft's equivalent—bundled into Foundry effective January 2026—mirrors these capabilities using Entra for identity and managed memory at session, user, and procedural scopes. Google, having retired the Vertex AI branding, now houses equivalent features under its Gemini Enterprise Agent Platform with deployments, memory banks, sessions, agent registries, and policy controls.
The architectural convergence reflects rational self-interest: infrastructure companies stack features vertically because integration generates margin. The cost lands on customers. Identity, logging, and deployment all terminate within one vendor's ecosystem, making the operational substrate beneath the agent extraordinarily difficult to relocate.
What Portability Would Require#
A foundational agent contract—modeled after what Cloud Foundry accomplished—would need to address three core gaps:
