Gemini Enterprise Agent Platform vs Bedrock AgentCore vs Azure AI Foundry

By Camellia · Updated Nov 3, 2026

If you run most of your workloads on one hyperscaler, the "bedrock agentcore vs azure ai foundry" question usually arrives with a third name attached: Google's agent platform. All three clouds now sell the same idea. You write an agent in a framework you like, and the cloud provides the hosting, the agent's identity, its memory, a managed path to tools, tracing and a place to enforce policy. What differs is how much each one assumes you already live inside its cloud, and how each one bills the pieces.

This is written for the CTO, head of platform or cloud architect who has to pick a landing zone for production agents, not a model leaderboard. Two naming notes first, because both changed. Google's platform is now called Gemini Enterprise Agent Platform. Microsoft's is Microsoft Foundry, with the agent hosting product named Foundry Agent Service, which many teams still search for as Azure AI Foundry. AWS's is Amazon Bedrock AgentCore. If you want the packaged, business-user side of these vendors instead, see Copilot Studio vs Gemini Enterprise. This page is the developer-platform layer underneath.

Key Facts

• Three platforms, one job. All three host agents, give them identities, store memory, expose tools and trace what happened. None of them requires you to use that cloud's own models.
• Different billing shapes. AWS bills AgentCore per service meter, such as $0.0895 per vCPU-hour and $0.00945 per GB-hour for v1 Runtime. Google bills one pool of Agent Compute at $0.085 per vCPU-hour and converts other services into vCPU-hours. Microsoft charges nothing extra to create or run Foundry-native agents, and bills container compute only for hosted agents.
• Identity is built in on all three. Foundry gives agents a dedicated Microsoft Entra identity, Google ships Agent Identity, and AgentCore Identity is compatible with existing identity providers such as Okta and Entra ID.
• Model choice is open. AgentCore works with any foundation model in or outside Bedrock, Google offers a Model Garden of first-party, third-party and open models, and Foundry swaps models from its catalog without changing agent code.
• The real lock-in is the neighbors. Entra, IAM, Google Cloud IAM, your data warehouse and your network design pull harder than any single agent API.

TL;DR

What Each Platform Is Built For

Gemini Enterprise Agent Platform is Google's answer for teams that want one place to build, deploy and govern agents next to Gemini models and Google Cloud data. Google lists Agent Runtime, Sessions, Memory Bank, Agent Gateway, Agent Identity, Agent Registry, Model Garden, observability and governance policies as the building blocks. Google's pitch is a single governed path rather than a menu of independent services.

Amazon Bedrock AgentCore is the most modular of the three. AWS describes modular services you can use together or independently: Harness, Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability, Payments, Evaluations, Optimization, Policy and Registry. You can adopt the Gateway in front of existing agents without moving them. That flexibility is the draw, and the bill has more meters to watch.

Microsoft Foundry is where the Azure estate meets the agent. Foundry Agent Service offers three agent types: prompt agents defined by configuration, voice-based prompt agents, and hosted agents where you bring your own code and framework as a container or a .zip file. The center of gravity is Microsoft Entra, Azure RBAC, virtual networks and publishing into Microsoft Teams and Microsoft Copilot.

If the question is which one ranks across the whole field including SaaS-native vendors, Best Enterprise AI Agent Platforms in 2026 and Best AI Agent Platforms in 2026 cover that wider shortlist.

Agent Runtime and Frameworks

The honest read: all three run containerized agent code, and all three have a lighter, config-first option in some form. Foundry is the one that most clearly separates "I want the least to manage" (prompt agents) from "I want full control" (hosted agents) in its own docs. AgentCore is the one that explicitly lets you use Runtime, Gateway or Memory alone.

Identity: Who Is the Agent?

Agents that call real systems need their own credentials, not a shared service account. This is where the three platforms line up with the identity provider you already run.

AgentCore is the most neutral here, because AWS says it works with identity providers you already have rather than assuming its own. Foundry is the most tightly coupled to one provider, which is a benefit if that provider is Entra and a constraint if it isn't.

Memory and Sessions

If memory is the deciding factor, our guide to agent memory tools compares standalone options that work across all three clouds.

Tools and Gateways

Agents become useful when they can reach your APIs. All three now expose a governed layer between the agent and the tool, and all three lean on the Model Context Protocol.

For a dedicated look at this layer, Best MCP Gateways in 2026 covers independent gateways that sit in front of any of these.

Observability and Evaluation

Because AgentCore emits OpenTelemetry, it is the easiest to pipe into a tool you already own. If you'd rather use an independent tracer, see Best LLM Observability Tools in 2026.

Model Choice

None of the three forces a single model family. The bigger question is where model tokens get billed: usually separately from the agent runtime, on the vendor's model pricing page.

Governance and Policy

AgentCore Policy stands out because it evaluates a rule before each tool call executes, which suits regulated workflows where an auditor wants a deterministic answer. Foundry's strength is organizational: it plugs into the RBAC model your security team already administers. For a full certification and compliance cut across vendors, use the enterprise platform roundup.

Pricing: Three Different Meters

These platforms don't share a unit, so a headline rate comparison misleads. Here is each vendor's own published meter.

Region matters. Google's agent compute rate has no region attached on its page, and its other tables use Iowa (us-central1). AWS's AgentCore table names no region for the rates above. The Azure figures are for East US and move by region. AgentCore also lists a v2 Runtime at $0.0169 per GB-hour for memory (consumption), so confirm which Runtime generation your quote assumes.

A Worked Example (arithmetic on the listed rates)

Take one hosted agent container with 1 vCPU and 2 GiB of memory running 100 hours in a month, ignoring model tokens, tools and storage. At Google's list rate that is 100 x ($0.085 + 2 x $0.009) = $10.30. On AgentCore's v1 Runtime rates it is 100 x ($0.0895 + 2 x $0.00945) = $10.84. On Foundry's East US hosted meters, assuming the memory rate applies per GiB-hour as the pricing page labels it, it is 100 x ($0.0994 + 2 x $0.0118) = $12.30. The gap is a couple of dollars. In practice the compute is small and the surrounding meters decide the bill: tool calls through a gateway, memory reads and writes, evaluations, and above all model tokens.

Two things to read carefully. Google says idle time waiting for the next prompt between turns is not billed, which can matter for chat-style agents. And Google's memory, sessions and registry reads and writes are converted into vCPU-hours, so a memory-heavy agent shows up as compute, not as a separate line.

Which Cloud Does Each Platform Pair With?

When Each One Is the Right Call

Gemini Enterprise Agent Platform fits when your data platform is already Google Cloud and you want identity, registry, gateway and memory in one governed product. The tradeoff is that several pieces bill through a shared vCPU-hour conversion, so you need usage data before the budget is predictable, and some pieces, such as Semantic Governance Policy, have billing that starts later.

Bedrock AgentCore fits when you want to take the parts you need. It's the strongest choice for teams that already have agents in LangGraph, CrewAI or Strands and want a gateway, identity and policy layer without a rewrite. The tradeoff is the number of meters: Runtime, Gateway, Identity, Memory, Policy, Evaluations and CloudWatch each bill on their own unit.

Microsoft Foundry fits when Entra is your identity backbone and the people consuming the agents work in Teams or Microsoft 365. Foundry-native agents carry no additional charge, which makes early pilots cheap. The tradeoff is that the Azure pricing page itself renders no numbers without the pricing calculator, so use the retail prices API or calculator for a real quote, and plan for regional differences.

Decision Framework

What to Do Next

Pick one production-shaped agent, not a demo. Log its real numbers for a week: hours of runtime, tool calls per task, memory reads and writes, and model tokens. Then price those numbers on each vendor's meters from the table above, in the region you will actually deploy. That exercise settles most of this decision, because the compute rates are close and the surrounding meters are not.

Then ask your security team one question that no feature table answers: which identity provider and policy engine do they already operate? The platform that fits that answer will usually be easier to approve than the one with the longest feature list. If you are still deciding whether you need a cloud platform at all, start with Best AI Agent Platforms in 2026 and Best Copilot Studio Alternatives.