Big Three Agent Platforms: AWS vs Microsoft vs Google — June 2026

We ran a three-lane research sweep on the big-three cloud agent platforms. This is the comparison, with every number cited. If you’re choosing what to build your agents on in 2026, the headline is uncomfortable: the platforms are converging on the parts that used to be differentiators, and the real decision is about lock-in and economics, not features.

For the open-source and personal-harness side of this same landscape, see the companion post: State of the Agent Harness — June 2026.

#AWS: agent infrastructure as a product category

AWS’s bet is that agents need infrastructure the way web apps needed EC2, and that AWS should sell that infrastructure by the component.

Bedrock AgentCore went GA on 2025-10-13 with all seven components at once: a serverless Runtime (session sandbox, 8-hour max, billed on active CPU so I/O wait is free), Memory (short-term events plus long-term semantic and episodic memory), a Gateway that turns REST and Lambda into MCP tools with semantic tool search, Identity (OAuth2/IdPs), Browser, Code Interpreter, and Observability (OpenTelemetry). December 2025 added Policy — natural-language policy compiled to deterministic enforcement outside the LLM loop — and Evaluations. April 2026 brought an AgentCore Payments layer (Coinbase x402 + Stripe Privy, USDC on Base, ~200ms settlement, per-session spend limits) and an Agent Registry — note: a private enterprise catalog, not a public marketplace, that indexes agents across AWS, Azure, GCP, and on-prem.

Pricing is pure infrastructure metering: Runtime/Browser/Code Interpreter at $0.0895 per vCPU-hour plus$0.00945 per GB-hour (active time only), Gateway at $0.005 per 1,000 tool calls, Memory at$0.25 per 1,000 short-term events. The honest caveat AWS won’t print: at sustained high throughput, self-hosting wins on cost.

For multi-agent orchestration, Bedrock shipped supervisor + sub-agent collaboration GA in March 2025, and the open-source Strands Agents SDK (Apache 2.0, v1.0 July 2025) gives Agents-as-Tools, Handoffs, Swarms, and Graphs across 20+ model providers — including non-AWS ones.

Two more AWS moves matter. Nova Act, the browser agent, went GA on 2025-12-02 and invented a pricing unit nobody else had: **$4.75 per agent-hour**, elapsed time, not tokens. And **[Amazon Kiro replaces Amazon Q Developer](https://aws.amazon.com/blogs/devops/amazon-q-developer-end-of-support-announcement/)** — Kiro launched internationally on 2026-05-07, Q blocked new signups on 2026-05-15, and Q hits full end-of-support on 2027-04-30. Kiro is a spec-driven IDE (requirements.md / design.md / tasks.md, hooks, steering files, custom subagents); [its pricing](https://kiro.dev/pricing/) runs Free / Pro$20 / Pro+ $40 / Power$200 on a credit model.

#Microsoft: one SDK, $0 runtime, 300M seats of distribution

Microsoft’s bet is the opposite of AWS’s. Give the orchestration layer away, charge for the tokens and tools that flow through it, and use the Microsoft 365 install base as a distribution moat no one else can match.

Microsoft Agent Framework hit v1.0 GA on 2026-04-03 as the successor to both AutoGen and Semantic Kernel (now maintenance-only). It’s MIT-licensed, .NET and Python, with graph-based Workflows (checkpointing, human-in-the-loop), native MCP and A2A, and multi-provider support including Anthropic and Ollama. If you built on AutoGen or SK, you’re migrating.

The pricing tell is Foundry Agent Service: the runtime itself is $0 — you pay tokens plus tool calls. Hosted Agents go GA around July 2026, framework-agnostic (Agent Framework, Copilot SDK, LangGraph) with session isolation. The point is to make running agents free so the consumption meter does the earning.

Distribution is where Microsoft is uncatchable. Copilot Studio got computer-using agents (UI automation) GA in May 2026, and the Agent Store in M365 Copilot has four intake paths into a 300M+ seat distribution surface. The new Work IQ APIs (GA 2026-06-16) expose M365 org intelligence — email, calendar, meetings, files, people — as 10 generic MCP tools billed on Copilot Credits.

Then there’s the part no other OS vendor is matching: Agentic Windows. MAI models (Nano 1.8B local / Core 7B / Pro 70B), a Windows Agent Runtime, Microsoft Execution Containers with Hyper-V isolation and hardware attestation, and Microsoft IQ — a cryptographically signed audit record of every agent action. Microsoft is trying to make Windows itself the agent substrate. And on the developer surface, GitHub is the front door: the Copilot coding agent (issue → PR via Actions), Agent HQ as a multi-vendor control plane (Anthropic, OpenAI, Google, Cognition, xAI all as GitHub Apps), and Agentic Workflows where a Markdown file in .github/workflows/ describes a goal and can only write through pre-approved safe outputs — never auto-merge.

#Google: one rebrand, the protocol, and the edge

Google’s bet has three legs: consolidate everything into one enterprise platform, own the inter-agent protocol as neutral standard, and push open-weight models down to the edge to lock in the local-agent tier.

Vertex AI was rebranded into the Gemini Enterprise Agent Platform (GA 2026-04-22 at Cloud Next), absorbing Agentspace. Agent Engine runtime bills at $0.0864 per vCPU-hour plus$0.0090 per GB-hour, with Sessions and Memory Bank GA at $0.25 per 1,000 stored events. The platform is **per-seat**:$21/user/month Standard, scaling up. Add Agent Studio (low-code) and Agent Garden (templates), and the open-source ADK (v2.2.0, 20k+ stars, Apache) deploying to Agent Engine, Cloud Run, GKE, or local.

The strongest Google play is A2A. The Agent-to-Agent protocol hit v1.0 stable on 2026-04-09, under the Linux Foundation since June 2025, with 150+ organizations (AWS, Microsoft, Cisco, IBM, Salesforce, SAP, ServiceNow), 22k+ stars, and native support inside both Bedrock AgentCore and Azure Foundry. Google authored the standard its competitors now run. AP2 (Agent Payments Protocol) extends the same playbook to agentic payments with 60+ orgs.

Google’s developer-tool sunset is the sharpest: Gemini CLI dies on 2026-06-18 for free, Pro, and Ultra tiers, pushing developers to the Antigravity CLI. Antigravity 2.0 (I/O 2026) is a VS Code-based desktop with a Cascade orchestration panel, four parallel specialist subagents, an MCP plugin schema, and competitor-settings importers. And Gemma 4 (Apache 2.0) is the edge play: the 12B encoder-free unified multimodal model (256K context, runs on a 16GB laptop, GPQA 78.8) plus E2B/E4B variants for phones and Raspberry Pi — a bid to own the local agent runtime tier.

#Where the three converge

Strip the branding and three convergences jump out:

• They run each other. GPT-5.5 and GPT-5.4 Codex are now on Bedrock; A2A runs natively in both Azure and AWS; Claude runs in Microsoft Agent Framework. The protocol layer — MCP plus A2A, both Linux Foundation — is neutral ground now, not anyone’s moat.
• They all want agents to spend money. All three shipped agentic payment rails in 2026 — AWS AgentCore Payments (x402/USDC), Google AP2, Microsoft via Marketplace billing. They expect agents to be economic actors within a year or two.
• They all sunset a dev surface in a 12-month window. Q Developer → Kiro, Gemini CLI → Antigravity, AutoGen/SK → Agent Framework. Building on big-three developer tooling carries migration risk by default — that’s three forced migrations in one year across three vendors.

Where they diverge is unit economics, and it’s worth saying plainly: AWS bills infrastructure (active CPU), Microsoft bills consumption (tokens + credits, $0 runtime), Google bills **per-seat plus infra**. Nova Act even invented elapsed-time agent pricing at$4.75/hour. Same category, three different meters — and your bill depends entirely on whether your agents are CPU-heavy, token-heavy, or seat-heavy.

#Where Mumega sits

Here’s the honest part. On managed-plane features, the big three are ahead of us and will stay ahead — they have AgentCore’s policy engine, Agentic Windows’ hardware attestation, A2A’s 150-org network. We don’t out-feature a hyperscaler and we don’t try to.

But all three share one shape: they’re managed planes. You rent agents that run on the vendor’s infrastructure, metered by the vendor, governed by the vendor’s control plane. None of them ships a sovereign agent substrate — a per-tenant agent runtime on your box, with your model keys, that you could be handed the keys to and run without depending on anyone’s plane. That lane is empty, and it’s the one we’re in.

A Mumega tenant gets a brain daemon — a perceive → think → decide loop on a hard token budget — running natively on their own infrastructure, reading their datasources, posting work to their GitHub, coordinating a colony through multi-agent orchestration over a bus with scoped tokens. The portability trick is keeping authorization in the application layer (sovereign, cloud-agnostic) and delegating only the authentication perimeter to whatever cloud the tenant chose. Multiple tenant brains run live this way today.

If you want managed convenience and you’re already all-in on one cloud, the big three are a reasonable bet — just price the migration risk in. If you need agents that run on infrastructure you own, with credentials you control, that survive whether or not a vendor keeps the product alive — that’s the substrate none of them sells, and the reason we built one. For the open-source harness side of the same map, read the companion: State of the Agent Harness — June 2026.

#Sources

• AWS Bedrock AgentCore — https://aws.amazon.com/bedrock/agentcore/
• AgentCore Payments (Coinbase + Stripe) — https://aws.amazon.com/blogs/machine-learning/agents-that-transact-introducing-amazon-bedrock-agentcore-payments-built-with-coinbase-and-stripe/
• AgentCore pricing — https://aws.amazon.com/bedrock/agentcore/pricing/
• AgentCore pricing vs. self-hosting — https://scalevise.com/resources/agentcore-bedrock-pricing-self-hosting/
• Bedrock multi-agent collaboration — https://aws.amazon.com/about-aws/whats-new/2025/03/amazon-bedrock-multi-agent-collaboration/
• Amazon Q Developer end-of-support / Kiro — https://aws.amazon.com/blogs/devops/amazon-q-developer-end-of-support-announcement/
• Kiro pricing — https://kiro.dev/pricing/
• Strands Agents SDK — https://github.com/strands-agents/sdk-python
• Nova Act GA — https://aws.amazon.com/blogs/aws/build-reliable-ai-agents-for-ui-workflow-automation-with-amazon-nova-act-now-generally-available/
• OpenAI models on Bedrock — https://www.aboutamazon.com/news/aws/bedrock-openai-models
• Microsoft Agent Framework overview — https://learn.microsoft.com/en-us/agent-framework/overview/
• Foundry Agent Service (Build 2026) — https://devblogs.microsoft.com/foundry/agent-service-build2026/
• Microsoft 365 Copilot Agent Store — https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-agent-store
• Work IQ APIs — https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/02/announcing-the-new-work-iq-apis/
• Microsoft Build 2026 (Agentic Windows) — https://news.microsoft.com/build-2026/
• GitHub Agent HQ / Agentic Workflows — https://github.blog/news-insights/company-news/welcome-home-agents/
• Gemini Enterprise Agent Platform — https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-agent-platform
• Google ADK — https://github.com/google/adk-python
• A2A protocol surpasses 150 organizations — https://www.linuxfoundation.org/press/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year
• MCP/A2A/ACP protocol convergence — https://zylos.ai/research/2026-03-26-agent-interoperability-protocols-mcp-a2a-acp-convergence
• Gemini CLI → Antigravity CLI transition — https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/
• Gemini CLI shutdown coverage — https://www.theregister.com/ai-ml/2026/05/20/bye-bye-gemini-cli-google-nudges-devs-toward-antigravity/5243605
• Gemma 4 12B — https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemma-4-12b/