Amazon, Microsoft, and Google are converging on the same enterprise agent architecture
The New Stack Janakiram MSV ● Covered by 3 sources
Amazon, Microsoft, and Google have all built nearly identical enterprise AI agent platforms, just with different names. That means your agent's identity, memory, and logs get locked into whichever cloud you pick.
Based on reporting by The New Stack, Janakiram MSV — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Something quietly clicked into place over the last nine months. Amazon shipped Bedrock AgentCore. Microsoft renamed Azure AI Foundry to Microsoft Foundry. Google folded Vertex AI into the Gemini Enterprise Agent Platform. Different names, different marketing decks, but strip away the branding and you find the same seven-ish pieces everywhere: a runtime, memory, a tool gateway, identity, observability, governance. Eighteen months ago this space was a scattered pile of libraries. Now it looks like a platform layer, and all three hyperscalers landed on the same shape without coordinating.
The article draws a sharp comparison to platform-as-a-service history. Back between 2011 and 2016, developers wired together VMs, load balancers, message queues, and monitoring tools by hand, each with its own quirks. Cloud Foundry and Heroku fixed that by giving applications a portable contract — buildpacks detected the language, service brokers wired up databases, and developers stopped caring which machine anything ran on. Cloud Foundry itself didn't end up winning the platform war, Kubernetes did, but the ideas survived and got rebuilt on top of Kubernetes through projects like Korifi.
Agents don't have that contract yet. AgentCore offers eight-hour execution windows and session isolation, with a gateway that turns Lambda functions and APIs into tools agents can call, and observability piped through OpenTelemetry into CloudWatch. Microsoft's Foundry Agent Service does the session-isolated runtime thing too, with Entra Agent ID for identity and memory split across session, user, and procedural scopes. Google's Gemini Enterprise Agent Platform renamed Agent Engine to Deployments and stacked Memory Bank, Sessions, Agent Registry, Policies, and Gateways around it. Same primitives, three different walled gardens — and once your agent's identity, telemetry, and state all live in one cloud's systems, moving it a year later means rebuilding the whole thing.
The pieces for something more open already exist, just not assembled. The Model Context Protocol handles tool access, A2A covers agent-to-agent communication, and OpenTelemetry is working out GenAI conventions for spans and token usage, though a lot of that is still in development. The Linux Foundation launched the Agentic AI Foundation in December 2025, with MCP, goose, and AGENTS.md as founding projects, and AWS, Google, and Microsoft all joined as platinum members. Google even moved A2A under the Linux Foundation's roof. But protocols aren't a lifecycle. Nothing in that foundation today tells you how to version an agent, promote it through environments, or roll it back after a bad evaluation.
Agents also aren't just web apps with a model bolted on, which is part of why copying the old PaaS playbook exactly won't work. Behavior is probabilistic, so identical inputs can trigger different tool calls. Agents act with a user's delegated authority, so a permissions bug becomes a real-world incident rather than a broken page. And an agent's behavior can shift without any code changing at all, just from a model update or a tweaked tool description. LangGraph shows the open-source version of handling this, with checkpointing and crash recovery, but the commercial control plane around it — deployment, evaluation, observability — sits in LangSmith, which shows fragmentation even inside a single project.
Nobody has answered the three questions that would actually matter to an enterprise doing diligence: who governs the project, whether the same agent artifact runs unmodified on two clouds, and whether memory lives somewhere exportable. Right now the answer to all three is no, everywhere. Kubernetes eventually gave the industry a shared vocabulary for running software; agents haven't gotten there, and whoever defines that vocabulary decides what an agent even is.
My take — AI-written commentary, not fact-checked reporting
Nobody should be surprised three hyperscalers converged on identical architecture and called it innovation — that's just what happens when the profit motive points at vertical integration. The interesting tell is that AWS, Google, and Microsoft all joined the Agentic AI Foundation as platinum members while still building proprietary lock-in on top of the same open protocols; that's not generosity, that's covering the exits. Enterprises betting big on any single vendor's agent stack right now are re-learning a lesson the cloud-native world already paid tuition for once.
Read more about this at: The New Stack