AWS Bedrock AgentCore vs Google Gemini Enterprise Agent Platform

AWS and Google both rebuilt their agent offerings in 2026. AWS expanded Amazon Bedrock AgentCore with Policy, Evaluations, and a managed Harness, and Google launched Gemini Enterprise Agent Platform as the evolution of Vertex AI. Engineering teams choosing where to build production agents usually end up comparing the two directly.

AgentCore vs Gemini Enterprise Agent Platform at a glance

AWS Bedrock AgentCore is a set of modular services for building, deploying, and operating agents with any framework and any model. The core pieces are Harness (a managed agent loop defined by configuration), Runtime (a dedicated microVM per session), Memory, Gateway (APIs and Lambda functions exposed as MCP tools), Identity, Code Interpreter, Browser, Observability, Evaluations, and Policy, with Registry, Optimization, and agent Payments alongside. Each service can be used on its own.

Google Gemini Enterprise Agent Platform was announced at Google Cloud Next on April 22, 2026, and Google says all Vertex AI services will be delivered through it going forward. It is organized into Build (the open-source Agent Development Kit, low-code Agent Studio, Agent Garden templates, Model Garden), Scale (Agent Runtime, Agent Sandbox, Memory Bank, Sessions), Govern (Agent Identity, Agent Registry, Agent Gateway with Model Armor, threat detection), and Optimize (simulation, evaluation, observability, and an optimizer).

Building agents

AgentCore. Developers work through the AgentCore CLI, SDKs, and APIs. Harness, generally available since June 2026, lets a team define an agent with a model, instructions, and tools and run it without writing orchestration code. For custom logic, Runtime hosts agents built with LangGraph, CrewAI, LlamaIndex, Strands, OpenAI Agents SDK, Google ADK, or your own code.

Gemini Enterprise Agent Platform. ADK is open source in Python, TypeScript, Go, Java, and Kotlin, and Google says it can be containerized and run anywhere. Agent Studio adds a visual, low-code path that exports to ADK, and Agent Garden provides templates, including financial analysis and invoice processing. This gives Google a stronger story for less technical builders.

Runtime, memory, and security

AgentCore. Every session gets its own microVM with isolated compute, memory, and filesystem, terminated and sanitized when the session ends. Sessions run up to 8 hours on microVMs or up to 14 days on instances. Policy, generally available since March 2026, intercepts every tool call at the Gateway and checks it against rules written in natural language or a Cedar-compatible language. Identity works with Okta, Entra ID, Cognito, and other providers.

Gemini Enterprise Agent Platform. Agent Runtime advertises sub-second cold starts and supports long-running agents that work for days. Memory Bank generates long-term memories from conversations. Agent Identity gives each agent a cryptographic ID with auditable authorization policies, and Agent Gateway applies consistent security policies and Model Armor protections against prompt injection and data leakage. Security Command Center powers an agent security dashboard.

Models and ecosystem

AgentCore. AWS says Runtime works with any foundation model in or outside Bedrock, including OpenAI, Gemini, Claude, Nova, Llama, and Mistral, and Harness supports Bedrock, OpenAI, Gemini, and OpenAI-compatible providers. Observability emits OpenTelemetry into CloudWatch.

Gemini Enterprise Agent Platform. Model Garden lists more than 200 models, with Gemini first-party and Claude Opus, Sonnet, and Haiku among third-party options. Event-driven and batch agents connect natively to BigQuery and Pub/Sub. Google names PayPal and Payhawk among its agent customers in financial services.

Pricing

AgentCore is consumption-based with no upfront commitment. Runtime microVMs start at $0.0895 per vCPU-hour and $0.00945 per GB-hour, billed per second on active use. Gateway is $0.005 per 1,000 invocations, and Memory, Policy, Evaluations, Browser, and Code Interpreter are metered separately.

Gemini Enterprise Agent Platform is also consumption-based. Agent Runtime and sandbox compute are $0.085 per vCPU-hour plus memory per GiB-hour, idle time between turns is not billed, and Memory Bank, Sessions, and Gateway usage are metered. On both platforms, model usage is billed separately.

Where Runtime fits

Both platforms give an engineering team excellent parts. Neither is the finished product an operations team uses. Building on either still means designing which payment actions need a human, building the approval flow and the interface, wiring in your ledger, processor, and bank portals, and producing an audit export a sponsor bank will accept. An engineering team can spend two quarters on that before the first agent touches real data.

Runtime is the AI agent harness for payment and fintech teams. Payment ops, risk, compliance, finance, and support build agents from their SOPs and call them from Slack, Teams, email, or the dashboard. Agents start read-only, stop for approval before releasing a payout, applying a reserve, or booking a ledger entry, and every run is recorded end to end. Runtime runs agent computers in your AWS or GCP account (or Azure, or fully self-hosted), so it is not a choice against either cloud. It routes across harnesses and models with fallbacks, and comes with a forward-deployed AI engineer who has built payment infrastructure.

Which should you choose

• Choose AgentCore if you run on AWS, want to adopt agent services one at a time, and value a deterministic policy engine on every tool call.
• Choose Gemini Enterprise Agent Platform if you run on Google Cloud, want Gemini first-party, or want a low-code studio next to a code-first framework.
• Choose Runtime if the goal is to put payment ops, risk, and finance teams on agents with approvals and an audit trail built for money movement, across whichever clouds you already use.