← Field Notes

Field note · AWS and Google Cloud · July 2026

Amazon Bedrock AgentCore vs Gemini Enterprise Agent Platform

AWS offers modular AgentCore services. Google Cloud now organizes agent development under Gemini Enterprise Agent Platform across Build, Scale, Govern and Optimize. The architectures overlap, but their centers of gravity differ.

AudienceEnterprise architects · CTOs · AI platform and data leaders
PerspectiveAWS and Google Cloud agent platform selection
Reading time16 minutes
AWSAmazon Bedrock AgentCore

Modular agent runtime and operating services.

ComparisonClosest capabilities, not identical products

Compare responsibilities, maturity and trust boundaries.

AlternativeGemini Enterprise Agent Platform

Integrated agent and enterprise governance platform.

Use the current Google Cloud product boundary

Google Cloud has evolved Vertex AI agent capabilities into Gemini Enterprise Agent Platform. The platform is organized around Build, Scale, Govern and Optimize, while the Gemini Enterprise app is the employee-facing place to discover, share and run agents.

Older Vertex AI Agent Engine documentation remains useful for migration context, but new architecture should use the current Agent Platform Runtime, ADK, Agent Identity, Agent Gateway, Memory Bank, evaluation and observability vocabulary.

Build and runtime

AgentCore Harness provides a managed agent loop, while Runtime hosts custom framework code. Google Agent Development Kit is a modular model-agnostic framework, Agent Studio provides a visual build experience, and Agent Platform Runtime deploys ADK, A2A, LangChain and custom agents.

Google’s abstraction begins with an application-development platform and framework ecosystem. AWS begins with modular runtime and operating services that are framework independent.

Tools, MCP and A2A

AgentCore Gateway fronts tools, agents and models, supporting MCP, A2A and multiple target types. Google Agent Platform supports a remote MCP server, Agent Gateway and Agent2Agent deployment through Agent Platform Runtime.

Both clouds are embracing open protocols. The architectural value comes from adding identity, policy, discovery, rate limits, evidence and lifecycle around those protocols. MCP and A2A alone do not create a production control plane.

Agent identity and runtime protection

AgentCore Identity gives AWS agents workload identities and delegated credentials, while AgentCore Policy authorizes tool calls through Gateway.

Google Agent Identity uses strongly attested SPIFFE-based identities with cryptographically bound credentials. Agent Gateway and Model Armor protect interactions and enforce runtime controls. The mechanisms differ, but the design principle is the same: every agent needs a first-class identity and policy must remain outside model reasoning.

Sessions, memory and code execution

AgentCore Memory supports short-term and long-term memory, and AgentCore Code Interpreter supplies isolated code execution. Google Agent Platform Sessions stores interaction state, Memory Bank generates and retrieves durable memories, and Agent Platform Code Execution runs code in a managed sandbox.

Do not treat the availability of a memory service as permission to store everything. Define memory topics, retention, deletion, tenant isolation and provenance before enabling automatic persistence.

Observability, evaluation and optimization

AgentCore provides CloudWatch-based observability, trace and span evaluations, custom evaluators, online measurement and configuration optimization.

Google Agent Platform exposes traces, topology visualization, offline evaluation, simulated evaluation, online monitoring and prompt optimization. Google’s topology view is particularly aligned with multi-agent graph operations, while AWS’s modular evaluation services integrate naturally with CloudWatch and Gateway traffic.

Data and model ecosystems

AWS combines Amazon Bedrock models, Managed Knowledge Bases, Data Automation and open framework support. Google Agent Platform combines Gemini, partner and open-weight models, Model Garden, RAG Engine and Google data grounding.

The model catalog should not decide the platform alone. Evaluate data gravity, identity, private networking, operational skills, regional availability and the cost of moving evidence through the system.

Closest service map

ResponsibilityAWSGoogle Cloud
Framework / managed loopAgentCore Harness or any frameworkADK and Agent Studio
Custom runtimeAgentCore RuntimeAgent Platform Runtime
GatewayAgentCore GatewayAgent Gateway and remote MCP server
Agent identityAgentCore IdentityAgent Identity using SPIFFE-based attestation
Memory and sessionsAgentCore MemoryAgent Platform Sessions and Memory Bank
Code executionAgentCore Code InterpreterAgent Platform Code Execution
Evaluation and optimizationAgentCore Evaluations and OptimizationAgent evaluation, online monitoring and prompt optimization

Selection guide

Choose AWS when your operational systems and security controls are AWS-native and you want modular agent primitives that attach to many frameworks. Choose Google Cloud when Gemini, ADK, A2A, Google data services and a unified Build-Scale-Govern-Optimize platform are strategic advantages.

Confirm feature status, region support and data-governance requirements before committing. Product names and preview status are changing quickly in 2026.

Executive takeaway

AWS and Google Cloud now cover the same production lifecycle: build, run, connect, identify, remember, observe, evaluate and improve. The best architecture uses open contracts at the application boundary and native controls inside the chosen cloud.

Primary reading

  1. Gemini Enterprise Agent Platform overview
  2. Build with Gemini Enterprise Agent Platform
  3. Agent Platform Runtime
  4. Agent Identity overview
  5. Optimize your agents
  6. Gemini Enterprise Agent Platform APIs
  7. Google Cloud announcement
  8. Amazon Bedrock AgentCore overview