Google Gemini Enterprise Agent Platform vs Amazon Bedrock AgentCore
Independent research, updated November 9, 2026. Based on vendors' official pages and documentation. We have not tested these products hands-on.
Answer in brief: Both are hyperscaler agent clouds for engineering teams, so the deciding factor is usually the cloud you already run on. Google Gemini Enterprise Agent Platform (the renamed Vertex AI) covers the whole lifecycle in one place: low-code Agent Studio, the code-first ADK, a managed Agent Runtime, 200+ models in Model Garden, and Gemini Enterprise to deliver agents to employees. Amazon Bedrock AgentCore is a modular set of production services (Runtime, Gateway, Identity, Policy, Memory, Browser, Code Interpreter, Web Search). It suits teams that already have a framework and want secure infrastructure with per-second billing and no minimum fees.
Top pick: Google Gemini Enterprise Agent Platform for teams that want build, run, govern and employee rollout from one vendor. Runtime usage is billed to the nearest second, and idle time between prompts isn't billed. Visit Google Gemini Enterprise Agent Platform →
Choose Amazon Bedrock AgentCore instead if you're on AWS and want to bring your own framework and adopt services one at a time. Visit Amazon Bedrock AgentCore →
Where they sit on the market map
Hyperscaler agent clouds host the agent logic you write, often with open-source frameworks, and add identity, tool governance, memory and observability around it. Neither is a no-code suite like Copilot Studio. Google comes closer, through Agent Studio and the no-code agent builder in Gemini Enterprise.
Head-to-head at a glance
| Google Gemini Enterprise Agent Platform | Amazon Bedrock AgentCore | |
|---|---|---|
| Build | Agent Studio (low-code), ADK (code), Managed Agents API (Pre-GA) | Bring your own framework: LangChain, OpenAI Agents SDK, Claude Agent SDK, Strands or custom |
| Runtime | Agent Runtime (formerly Agent Engine); ADK integrated; managed templates for LangChain, LangGraph, AG2, LlamaIndex (preview) | Serverless microVMs, or Instances on EC2 in your own account (sessions persist up to 14 days) |
| Identity and policy | Agent Identity (SPIFFE format), Agent Registry, Agent Gateway, Model Armor | AgentCore Identity, Gateway access control, Policy with Cedar rules |
| Memory | Sessions and Memory Bank | AgentCore Memory |
| Models | 200+ in Model Garden, including Gemini, Anthropic Claude and Gemma | Any model |
| Interop | Agent2Agent (A2A) on Agent Runtime (preview); Registry stores MCP servers | Gateway reaches MCP servers, APIs and Lambda functions |
| Compute price | $0.085/vCPU-hour, $0.009/GiB-hour | Per-second CPU and memory used; CPU scales to zero during I/O wait |
| Employee front end | Gemini Enterprise (Business from $21/seat/month) | Not part of AgentCore |
Google: the complete lifecycle stack
Google's April 22, 2026 release notes renamed Vertex AI Agent Engine to Agent Runtime, and Agent Engine Sessions and Memory Bank to Agent Platform Sessions and Memory Bank. Google's documentation describes three build paths: Agent Studio for low-code visual work, ADK for custom code, and the REST-first Managed Agents API. The Managed Agents API is Pre-GA, and Google says it is not for production use during the preview. On governance, Agent Registry is the central store for agents and MCP servers, each agent can get its own SPIFFE-format Agent Identity, and Agent Gateway enforces access policies and Model Armor inspection on agent traffic. Distribution is the standout: Gemini Enterprise delivers agents to employees in Business, Standard, Plus, Pay-as-you-go and Frontline editions.
AWS: modular infrastructure, priced for how agents behave
AWS calls AgentCore the platform for production agents: any framework, any model, secure at scale. Its pricing is designed around how agents actually run. On microVMs, billing is per second with a 1-second minimum, and CPU drops to zero while the agent waits on an LLM, a tool or a database. Runtime v2 reclaims idle memory after 120 seconds. AgentCore Identity costs nothing extra when used through Runtime or Gateway. Policy checks every Gateway tool call against your rules, and you can write rules in natural language that are converted into Cedar policy statements. Each feature can be used on its own.
Pricing model compared
- Google: Agent Runtime is $0.085/vCPU-hour and $0.009/GiB-hour, rounded to the nearest second, with 1-year and 3-year savings plans. Sessions and Memory Bank each cost $0.30 per GiB-month of storage, $0.085 per 3 million reads and $0.085 per 1 million writes. Model and embedding tokens are billed separately.
- AWS: Runtime, Browser and Code Interpreter are billed by CPU and memory consumed, with a 128MB memory minimum. Instances cost the EC2 price plus a management fee, which is a percentage of the On-Demand price. Web Search is $7 per 1,000 queries and Web Answer is $10 per 1,000 tool calls. Gateway egress to customer VPCs is $0.006/GB. Models, storage and data transfer are billed separately. New AWS customers get up to $200 in Free Tier credits.
When to pick which
- Google: you're on Google Cloud, you want low-code and code-first building on one platform, you need broad model choice including Claude, or you plan to roll agents out to staff through Gemini Enterprise.
- AWS: you're on AWS, you already have an agent framework, you need long sessions on your own EC2 capacity (with your Savings Plans applied to the compute), or you want to adopt a single service such as Gateway or Code Interpreter without moving everything.
FAQ
Is Vertex AI still available?
Vertex AI is now Gemini Enterprise Agent Platform. Google's April 2026 release notes map the old names to the new ones, for example Agent Engine to Agent Runtime.
Can AgentCore run agents built with LangGraph or other frameworks?
AWS says AgentCore works with any framework. It specifically names LangChain, the OpenAI Agents SDK, the Claude Agent SDK and Strands, as well as your own framework.
Which one is cheaper?
It depends on the workload. Both bill compute and memory, and both bill model tokens separately. AgentCore doesn't charge CPU during I/O wait, and Google doesn't bill Runtime idle time between prompts. Model a real month of traffic on each.
Visit the platforms
Visit Google Gemini Enterprise Agent Platform → · Visit Amazon Bedrock AgentCore →
Keep going
Back to the ranked list · Google review · AgentCore review · How to choose. Comparing model hosting rather than agent platforms? See bestinferenceproviders.com.