The three big clouds now each have an enterprise agent offering, but they are not the same kind of product. Microsoft Agent 365 is first a governance and identity control plane - it watches and governs agents, including ones built elsewhere. AWS Bedrock AgentCore is code-first build infrastructure - modular runtime, memory, and tool gateways for engineers. Google Gemini Enterprise is a build-and-run platform grown out of Vertex AI, strong on persistent memory and data grounding. Pick by what you need most: govern a sprawling agent estate (Microsoft), build custom agents your way (AWS), or run data-heavy agents fast and Google-native (Google).
This is a teardown of tools we actually build on. No platform is "best" - they solve different problems, and the honest answer usually follows your existing cloud, your identity system, and who is building the agents. Here is how the three differ, and how to choose.
What changed
Through 2026 the three major clouds stopped shipping standalone agent toolkits and consolidated them into named enterprise platforms. Microsoft Agent 365 (announced in late 2025, generally available May 1, 2026) became the governance layer over the agents you build in Copilot Studio and Microsoft Foundry. AWS replaced its first-generation Bedrock Agents - now in maintenance mode - with Bedrock AgentCore, a decoupled, code-first stack. And Google folded its agent tooling out of Vertex AI into Gemini Enterprise (the rebrand landed April 22, 2026). Around all three, two open standards have become table stakes: the Model Context Protocol (MCP) for connecting agents to tools, and agent-to-agent (A2A) patterns for multi-agent coordination.
The three platforms, side by side
Microsoft Agent 365 - the governance and identity control plane. Microsoft's pitch is not "build agents here" so much as "govern every agent your company runs, wherever it was built." Its standout is identity: agents get a first-class Entra Agent ID, so security teams can apply conditional access, lifecycle governance, and auditing from the same directory that governs employees. Microsoft Defender adds shadow-agent discovery to flag unknown or rogue agents on the network (some of this shipped GA, some in preview). You still build agents in Copilot Studio (low-code) or Microsoft Foundry (full-code); Agent 365 is the layer that registers, monitors, and secures them - including agents running on other platforms, though enforcement is deepest inside the Microsoft estate.
AWS Bedrock AgentCore - code-first build infrastructure. AgentCore is the engineer's platform. Instead of a visual builder, it exposes modular primitives - Runtime, Memory, and Tool Gateways - and is deliberately framework-agnostic: custom agent logic written in LangGraph, CrewAI, or AutoGen drops in without a rewrite. Governance is AWS-native: IAM policies, Bedrock Guardrails, and tightly scoped tool gateways lock down every connection to a database, API, or service at the infrastructure layer. The trade-off is that there is almost no low-code path - AgentCore assumes a capable cloud-engineering team.
Google Gemini Enterprise - build-and-run, memory and data first. Google's platform pairs a low-code Agent Studio with an Agent Runtime tuned for fast cold starts, and its defining feature is Memory Bank - a persistent, cross-session memory layer that solves "agent amnesia" for long, multi-turn workflows. Governance comes through Agent Identity and an Agent Gateway, and the platform is strongest when agents are grounded in data that already lives in Google - BigQuery and Cloud Storage. The trade-off is that it is opinionated: you work inside Google's agent kits and structure.
| Microsoft Agent 365 | AWS Bedrock AgentCore | Google Gemini Enterprise | |
|---|---|---|---|
| Primary role | Governance + identity control plane | Code-first build infrastructure | Build-and-run platform |
| Who it is for | Security/IT governing an agent estate | Cloud engineers building custom agents | Data/ML teams, GCP-native devs |
| Build model | Copilot Studio (low-code) + Foundry (code) | Code-first; LangGraph/CrewAI/AutoGen drop in | Agent Studio (low-code) + Agent Runtime |
| Standout | Entra Agent ID + Defender shadow-agent discovery | Modular Runtime, Memory, Tool Gateways | Memory Bank + fast cold starts + data grounding |
| Governance | Enterprise directory (Entra), cross-platform | AWS IAM, Guardrails, tool gateways | Agent Identity + Agent Gateway |
Microsoft is selling you governance, AWS is selling you building blocks, and Google is selling you speed and memory. Buy the one whose strongest product is the problem you actually have.
How to actually choose
- You need to govern many agents across teams and tools. Microsoft Agent 365. If your pain is not building one agent but controlling a growing estate of them - identity, access, shadow-agent discovery - that is exactly what it leads with, and it reaches agents built elsewhere.
- You are an engineering-heavy shop building custom agents. AWS Bedrock AgentCore. If your team writes agent logic in its own frameworks and wants infrastructure-level control over every tool and permission, the modular, code-first model fits.
- Your advantage is data and you live on GCP. Google Gemini Enterprise. For data-analytics-heavy or long-running agents that need persistent memory and tight BigQuery grounding, Memory Bank and the fast runtime are the draw.
And the move that outlasts the choice: keep your agent logic, prompts, evals, and tool contracts as portable as you can. All three have converged on MCP and A2A precisely so agents are not welded to one platform - lean into that so the control plane stays replaceable. It is the same lesson from this week's Signal / Noise digest and from yesterday's teardown of OpenAI across Azure, AWS and Google Cloud: the durable work is the governance and portability around the agent, not the vendor logo on it.
Governance is not optional, either - it is what stands between a useful agent and the kind of incident where banks start blocking your agents because they cannot tell them apart from an attacker. If you are choosing where to build and govern AI agents, tell us your footprint and we will give you a straight read.