AI agent platforms are the managed services and open-source frameworks developers use to build, host and run agents that plan steps and call tools, and the five largest model and cloud providers each now offer one. This page compares the managed agent platforms of Amazon, Anthropic, Google, Microsoft and OpenAI, then seven widely used agent frameworks and SDKs, on identical fields: components, models, tool and protocol support, identity, observability, pricing basis, licence and lifecycle status. It is written for architects and developers choosing where to build agents and for buyers estimating cost. Platforms and frameworks are listed alphabetically; order is never a ranking. Every fact comes from the provider's or project's own documentation, cited in the sources list.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Confirm details against each provider's current documentation.
Key takeaways
- Five managed platforms, five billing models. Amazon Bedrock AgentCore bills per service consumed, Claude Managed Agents bills $0.08 per running session-hour plus tokens, Gemini Enterprise Agent Platform meters agent compute, memory and storage, Microsoft Foundry Agent Service bills inference, tools and container compute, and OpenAI states AgentKit tools are included in standard API model pricing.
- OpenAI Agent Builder shuts down on 30 November 2026. OpenAI's Agent Builder guide marks the visual builder as deprecated; ChatKit remains available and workflow code can be exported to run with the Agents SDK.
- MCP support is widely documented. All five managed platforms and five of the seven frameworks document Model Context Protocol tool connections in the sources checked; A2A support is documented by AgentCore, Gemini Enterprise Agent Platform, Foundry Agent Service, CrewAI and Strands Agents.
- Agent identity is now a platform feature. AgentCore Identity, Google's SPIFFE-formatted Agent Identity and Foundry's dedicated Microsoft Entra identity per hosted agent give agents their own identities instead of shared credentials.
- Most frameworks are permissively licensed. CrewAI, LangGraph, Microsoft Agent Framework and the OpenAI Agents SDK repositories carry MIT licences, ADK and Strands Agents carry Apache 2.0, and Claude Agent SDK use is governed by Anthropic's Commercial Terms of Service.
What are AI agent platforms and frameworks?
An agent framework is a library that runs the agent loop in your own process: it sends the model a prompt and tool definitions, executes the tools the model asks for and repeats until the task is done. A managed agent platform hosts that loop, or the agent's runtime, for you and adds services around it such as memory, sandboxes, identity, policy and tracing. The two layers overlap: Google states ADK agents deploy to Agent Runtime, Microsoft Agent Framework agents deploy to Foundry hosted agents, and AgentCore Runtime hosts agents written in CrewAI, LangGraph, LlamaIndex, ADK, the OpenAI Agents SDK and Strands Agents.
This page does not repeat the security trade-offs between the two layers, which are covered in Agent Frameworks vs Managed Agent Platforms. Low-code business builders are covered in the Low-Code Agent Platforms Guide, coding agents in Vibe Coding Tools Compared, consumer agents in Personal AI Agents, and the models themselves in Frontier AI Models Compared.
The managed platforms at a glance
As of 2026-11-15. Source: each provider's product page and documentation, listed in Standards and references.
| Platform | Provider | What the provider says it is | Status stated by provider | Main components named in documentation |
|---|---|---|---|---|
| Amazon Bedrock AgentCore | Amazon Web Services | A set of services for building, deploying and operating agents, usable together or separately | Status per service not stated on the overview page; Optimization Insights in public preview | Harness, Runtime, Memory, Gateway, Identity, Code Interpreter, Browser, Observability, Payments, Evaluations, Optimization, Policy, Registry |
| Claude Managed Agents | Anthropic | A pre-built, configurable agent harness that runs in managed infrastructure, for long-running and asynchronous tasks | Beta (header managed-agents-2026-04-01), enabled by default for all API accounts | Agent, environment (cloud or self-hosted sandbox), session, events |
| Gemini Enterprise Agent Platform | Google Cloud | A platform to build, scale, govern and optimise enterprise agents grounded in enterprise data; product page titled "Gemini platform" | Status per component not stated on the pages checked (2026-11-15) | Agent Development Kit, Agent Studio, Agent Runtime, Sessions, Memory Bank, sandboxes, Agent Garden, Agent Identity, Agent Gateway, Agent Registry |
| Microsoft Foundry Agent Service | Microsoft | A service to orchestrate and host AI agents to automate business processes | Prompt and hosted agents documented as available; voice-based prompt agents in preview | Prompt agents, voice-based prompt agents, hosted agents, Foundry IQ, tracing and evaluation |
| OpenAI AgentKit | OpenAI | A set of tools to build, deploy and optimise agents, built on the Responses API (announced 2025-10-06) | Agent Builder deprecated, shutting down 2026-11-30; ChatKit remains available | Agent Builder, ChatKit, Connector Registry, Evals, Guardrails |
Which models can each platform use?
| Platform | Models stated by provider |
|---|---|
| Amazon Bedrock AgentCore | Runtime: models in or outside Amazon Bedrock, including Claude, Nova, Llama, Mistral, OpenAI and Gemini. Harness: Amazon Bedrock, OpenAI, Google Gemini and any OpenAI-compatible provider |
| Claude Managed Agents | All Claude models, per the overview page |
| Gemini Enterprise Agent Platform | More than 200 Google and third-party models and tools in Model Garden, including Gemini 4 Argon, the Anthropic Claude family and open models such as Gemma |
| Microsoft Foundry Agent Service | Models from the Foundry model catalogue, such as GPT-4o, Llama and DeepSeek; models can be swapped without changing agent code |
| OpenAI AgentKit | OpenAI models through the Responses API; the launch post adds third-party model support in Evals |
For model specifications and prices, see the LLM Pricing Tracker.
What features does each platform include?
As of 2026-11-15. Each cell records what the provider's documentation describes.
| Platform | Hosted runtime and sandbox | Memory and state | Tools and MCP | Identity and access | Observability and evaluation |
|---|---|---|---|---|---|
| Amazon Bedrock AgentCore | Serverless Runtime with session isolation; Harness sessions in isolated microVMs with filesystem and shell; Code Interpreter (Python, JavaScript, TypeScript); managed Browser | Short-term and long-term memory; memory stores shareable across agents | Gateway converts APIs and Lambda functions into MCP tools and connects existing MCP servers; Registry catalogues agents, MCP servers and skills | Identity works with Amazon Cognito, Okta, Microsoft Entra ID and Auth0; Policy intercepts every tool call, with rules compatible with the open-source Cedar language | OpenTelemetry-based Observability with Amazon CloudWatch; Evaluations; Optimization |
| Claude Managed Agents | Anthropic-managed cloud sandbox, self-hosted sandbox, or the whole service self-hosted; Bash and file operations in the sandbox | Stateful sessions with persistent filesystems and conversation history | Built-in Bash, file operations, web search and web fetch with domain allow and block lists; MCP servers; skills | Credentials through vaults ("Authenticate with vaults" section); API key access | Events streamed over server-sent events; agents can be steered or interrupted mid-run |
| Gemini Enterprise Agent Platform | Agent Runtime hosting ADK or other agents; managed sandboxes for code execution, computer use and shell tasks | Sessions for single conversations; Memory Bank for long-term memory across sessions | MCP for tool connectivity; Agent Registry stores agents and MCP servers; Agent Garden sample library | Agent Identity issues a SPIFFE-formatted ID per agent instead of shared service accounts; Agent Gateway enforces access control and Model Armor inspection of tool calls | Cloud Observability with OpenTelemetry traces, logs and metrics; Gen AI evaluation service |
| Microsoft Foundry Agent Service | Prompt agents run with no infrastructure; hosted agents run your container or .zip code, each session in a VM-isolated sandbox connected to your virtual network | State persistence for hosted agents; Foundry IQ knowledge bases | Built-in tools, functions, OpenAPI specs and remote MCP servers | Microsoft Entra identity and managed identity; automatic dedicated Entra identity per hosted agent; OAuth on-behalf-of for tool connections | Tracing, metrics and evaluations with Application Insights; AI Red Teaming Agent (preview) |
| OpenAI AgentKit | Agent Builder workflows hosted by OpenAI until 2026-11-30; ChatKit embeds agent chat in your product | Agent Builder versioning: publishing creates a major version that API calls can pin | Connector Registry for pre-built connectors and third-party MCP servers; MCP node in Agent Builder | Connector Registry admin panel, which requires the Global Admin Console | Evals with datasets, trace grading and automated prompt optimisation; open-source Guardrails library |
Agent-to-agent protocol
AgentCore Runtime lists MCP and A2A as supported protocols. Google's agents documentation lists Agent2Agent. Foundry Agent Service documents an A2A endpoint for prompt agents, with A2A v1.0 generally available and v0.3 in preview, and A2A as a publishing channel. A2A support is not stated on the Claude Managed Agents overview or the AgentKit pages checked (2026-11-15). See the A2A Protocol Security Guide and the MCP Security Guide.
Where can you run each platform?
| Platform | Where it runs, as stated by provider |
|---|---|
| Amazon Bedrock AgentCore | Supported AWS Regions (listed on a separate regional availability page); Runtime Instances option runs on Amazon EC2; Registry covers resources on AWS, on-premises or other clouds |
| Claude Managed Agents | Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed); self-hosted sandbox or self-hosted service on your infrastructure |
| Gemini Enterprise Agent Platform | Google Cloud; ADK agents can also be containerised for Cloud Run, per the ADK repository |
| Microsoft Foundry Agent Service | Microsoft Foundry on Azure; bring-your-own virtual network; publishing to Microsoft Teams, Microsoft 365 Copilot, the Entra Agent Registry and A2A clients |
| OpenAI AgentKit | OpenAI platform; ChatKit and exported Agents SDK workflow code can run on your own infrastructure |
How much does each platform cost?
As of 2026-11-15. USD, list prices as stated on each provider's pricing page. Model tokens are billed separately on every platform unless the note says otherwise.
| Platform | Pricing basis | Example rates stated by provider | Free allowance stated by provider |
|---|---|---|---|
| Amazon Bedrock AgentCore | Consumption based, each service billed separately, no minimum fees | Runtime (microVMs v1) $0.0895 per vCPU-hour and $0.00945 per GB-hour; Gateway $0.005 per 1,000 invocations; Policy $0.000025 per authorisation request; Identity $0.010 per 1,000 tokens or keys for non-AWS resources (free through Runtime or Gateway). Region not labelled on the page | Up to $200 in AWS Free Tier credits for new customers; Registry monthly free tier |
| Claude Managed Agents | Session runtime plus tokens at standard model rates | $0.08 per session-hour while running (idle time not billed); web search $10 per 1,000 searches; Batch API discount does not apply | Not stated by Anthropic on the pages checked (2026-11-15) |
| Gemini Enterprise Agent Platform | Usage based; agent runtimes metered by vCPU, RAM and storage, plus model token pass-through | Agent Search about $4.00 per 1,000 queries; model tiers PayGo, Flex PayGo, Priority PayGo and Provisioned Throughput | Monthly per billing account: 50 hours of agent compute, 100 hours of agent RAM, 1 GiB of agent storage; $300 credit for new customers, valid 90 days |
| Microsoft Foundry Agent Service | Prompt agents: inference plus tool usage; hosted agents: inference, tool usage and container compute; voice agents: Voice Live and model usage | Rates not stated on the overview page checked (2026-11-15) | Not stated by Microsoft on the pages checked (2026-11-15) |
| OpenAI AgentKit | Included with standard API model pricing, per the launch post | No separate AgentKit fee stated | Not stated by OpenAI on the pages checked (2026-11-15) |
For cost guardrails on the credentials agents use, see How to set spend limits on LLM API keys.
Open-source agent frameworks and SDKs compared
As of 2026-11-15. Source: each project's documentation site and GitHub repository README and LICENSE files. Frameworks are listed alphabetically.
| Framework | Maintainer | Licence | Languages | Core abstractions | Managed hosting named by maintainer |
|---|---|---|---|---|---|
| Agent Development Kit (ADK) | Apache 2.0 | Python 3.11+; also Java, Kotlin, Go and TypeScript | Agents, Workflow Runtime (graph execution with routing, loops and retries), Task API for agent-to-agent delegation, Agent Config for no-code agents | Agent Runtime on Gemini Enterprise Agent Platform (formerly Vertex AI Agent Engine); Cloud Run | |
| Claude Agent SDK | Anthropic | Use governed by Anthropic's Commercial Terms of Service; Python repository carries an MIT LICENSE file | Python, TypeScript | The Claude Code agent loop as a library: built-in tools, hooks, subagents, permissions, sessions, skills, plugins | Runs in a process you operate; Claude Managed Agents is Anthropic's hosted option |
| CrewAI | CrewAI Inc | MIT | Python | Crews (role-based autonomous agents) and Flows (event-driven workflow control) | CrewAI AMP Suite, cloud or on-premise |
| LangGraph | LangChain Inc | MIT | Python; LangGraph.js for JavaScript and TypeScript | Low-level graph orchestration for long-running, stateful agents; Deep Agents built on top | LangSmith Deployment |
| Microsoft Agent Framework | Microsoft | MIT | Python, .NET; Go SDK in a separate repository | Agents, graph-based workflows (sequential, concurrent, handoff, group collaboration), middleware, declarative YAML agents, agent skills | Foundry hosted agents |
| OpenAI Agents SDK | OpenAI | MIT (Python repository) | Python; JavaScript and TypeScript SDK | Agents, handoffs and agents as tools, guardrails, sessions, tracing, sandbox agents, realtime agents | Runs on your infrastructure; exported Agent Builder workflows run on it |
| Strands Agents | Amazon Web Services | Apache | Python, TypeScript | Agents with multi-agent patterns: agents as tools, Swarm, Graph, Workflow | Amazon Bedrock AgentCore; also AWS Lambda, Fargate, App Runner, EKS and EC2 |
Models, protocols and controls in each framework
| Framework | Model providers stated | MCP and A2A | Human approval and durability | Observability |
|---|---|---|---|---|
| Agent Development Kit (ADK) | Optimised for Gemini, described as model-agnostic | MCP tools and OpenAPI specs; Task API for agent-to-agent delegation | Tool confirmation flow; human-in-the-loop in workflows | Evaluation built in; Cloud Observability when hosted on Agent Runtime |
| Claude Agent SDK | Claude models through API key authentication; claude.ai login not allowed for third-party products unless approved | MCP | Permissions control which tools run automatically or need approval; hooks at lifecycle points; sessions can be resumed or forked | Hooks can run custom code at lifecycle points |
| CrewAI | Various LLMs through several connection options; OpenAI API by default | MCP and A2A | Human input, checkpointing, structured outputs | Anonymous usage telemetry by default; tracing in the AMP control plane |
| LangGraph | Any model through LangChain integrations; usable without LangChain | Not stated in the LangGraph README checked (2026-11-15) | Durable execution that resumes after failure; interrupts for human inspection and edits | LangSmith tracing and evaluation |
| Microsoft Agent Framework | Microsoft Foundry, Azure OpenAI, OpenAI, GitHub Copilot SDK and others | Not stated in the Agent Framework README checked (2026-11-15) | Checkpointing, human-in-the-loop and time travel in workflows | Built-in OpenTelemetry integration |
| OpenAI Agents SDK | OpenAI models through the Responses API by default; LiteLLM and Any-LLM adapters for other providers | Remote MCP server tools | Human-in-the-loop interruptions with resumable state; sandbox agents with resumable sessions | Built-in tracing |
| Strands Agents | Amazon Bedrock, Anthropic, Google, LiteLLM, llama.cpp, Mistral AI, Ollama, OpenAI, Writer and others; custom providers | MCP (with MCP Router) and A2A | Agents run in your application process; deployment patterns for durable hosting | Metrics, traces and logs; Strands Evals toolkit |
Security guides for each framework: Google Agent Development Kit Security, Claude Agent SDK Security Guide, CrewAI security guide, LangChain and LangGraph Security Guide, Microsoft Agent Framework and AutoGen security guide, Securing OpenAI Agents SDK and Responses API Guide and AWS Strands Agents SDK Security Guide.
Lifecycle and naming changes
- Amazon: the AgentCore pricing page lists a committed baseline price for Runtime microVMs v2 "launching by October 2026" and Optimization Insights in public preview.
- Anthropic: Claude Managed Agents is in beta under the managed-agents-2026-04-01 header; MCP tunnels and dreaming are a more limited research preview requiring an access request. The Claude Agent SDK replaced the earlier Claude Code SDK packages, per its migration guide link.
- Google: the ADK repository states Agent Runtime was formerly Vertex AI Agent Engine, and the platform product page is titled "Gemini platform". ADK is at version 2.0, with a release cadence the repository describes as roughly bi-weekly.
- Microsoft: Microsoft Agent Framework documentation includes migration guides from AutoGen and Semantic Kernel. Voice-based prompt agents in Foundry are in preview without an SLA.
- OpenAI: the Agent Builder guide states Agent Builder is deprecated and shuts down on 30 November 2026; existing users can continue during the transition and ChatKit remains available.
For retirement planning across AI services, see the AI Model Retirement Guide.
Data handling and security
- Amazon Bedrock AgentCore: Policy enforces deterministic rules on every tool call before execution; Identity is free when used through Runtime or Gateway. See the AWS Bedrock AgentCore Security Guide.
- Claude Managed Agents: not eligible for zero data retention except when self-hosted, and not HIPAA eligible; transcripts persist until deleted, and sessions and files can be deleted through the API. See the Claude Agent SDK Security Guide.
- Gemini Enterprise Agent Platform: Agent Identity replaces shared service accounts, and Agent Gateway applies Model Armor inspection to tool calls. See the Vertex AI Agent Engine security guide and Google Cloud Agent Identity.
- Microsoft Foundry Agent Service: hosted agent sessions run in VM-isolated sandboxes connected to your virtual network, each hosted agent receives a dedicated Entra identity. See the Microsoft Foundry Agent Service Security Guide.
- OpenAI AgentKit: the Agent Builder guide points to prompt injection and data leakage risks and a separate safety guide; Guardrails provides PII masking and jailbreak detection. See the Securing OpenAI Agents SDK and Responses API Guide.
For the identity model behind agents, see the Agent Identity and Access Guide and Cloud Agent Identity Services Compared.
Benchmarks
None of the five providers or seven framework maintainers publishes a dated, comparable benchmark for its agent platform or framework on the pages checked (2026-11-15). Each platform's evaluation features are listed in the features tables above; model benchmarks are in Frontier AI Models Compared.
How to choose a platform: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Model choice | AgentCore, Gemini Enterprise Agent Platform and Foundry Agent Service document third-party models; Claude Managed Agents documents Claude models; AgentKit documents OpenAI models |
| Self-hosting | Claude Managed Agents offers self-hosted sandboxes and a self-hosted service; Foundry hosted agents run in your virtual network; every framework runs in your own process |
| Bring your own framework | AgentCore Runtime, Agent Runtime and Foundry hosted agents each name supported third-party frameworks |
| Per-agent identity | Documented for AgentCore, Gemini Enterprise Agent Platform and Foundry Agent Service |
| Billing shape | Per service consumed, per session-hour, per resource metered, per inference and compute, or included in model pricing |
| Product stability | Claude Managed Agents is beta; OpenAI Agent Builder ends 2026-11-30; Foundry voice agents are preview |
| Licence of framework code | MIT for CrewAI, LangGraph, Microsoft Agent Framework and OpenAI Agents SDK; Apache for ADK and Strands; Commercial Terms for Claude Agent SDK use |
Run a pilot on a representative task with your own evaluation set before committing to a platform.
Practitioner checklist
- Inventory every agent platform, framework and SDK version in use and record who owns each deployed agent.
- Migrate any OpenAI Agent Builder workflows before 30 November 2026 by exporting code or moving to ChatKit and the Agents SDK.
- Give each production agent its own identity where the platform supports it, rather than a shared service account or personal key.
- Enforce tool-call policy and human approval for consequential actions using the platform or framework controls listed above.
- Pin MCP server versions and tool lists, and review each MCP server before connecting it.
- Turn on tracing (OpenTelemetry where supported) and retain traces according to your logging policy.
- Check data retention and ZDR eligibility for the platform, not only for the model, before routing regulated data.
- Set budgets for both tokens and runtime charges, since several platforms bill compute or session time in addition to tokens.
Standards and references
- AWS: What is Amazon Bedrock AgentCore? (accessed 2026-11-15)
- AWS: Amazon Bedrock AgentCore pricing (accessed 2026-11-15)
- Anthropic: Claude Managed Agents overview (accessed 2026-11-15)
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Agent SDK overview (accessed 2026-11-15)
- Google Cloud: Gemini Enterprise Agent Platform agents documentation (accessed 2026-11-15)
- Google Cloud: Gemini platform product and pricing page (accessed 2026-11-15)
- Google: Agent Development Kit (adk-python) repository (accessed 2026-11-15)
- Microsoft Learn: Foundry Agent Service overview (accessed 2026-11-15)
- Microsoft: Agent Framework repository (accessed 2026-11-15)
- OpenAI: Introducing AgentKit (accessed 2026-11-15)
- OpenAI: Agent Builder guide (accessed 2026-11-15)
- OpenAI: Agents SDK documentation (accessed 2026-11-15)
- CrewAI: crewAI repository (accessed 2026-11-15)
- LangChain: LangGraph repository (accessed 2026-11-15)
- AWS: Strands Agents documentation (accessed 2026-11-15)
Related NHI Mgmt Group resources: Agent Frameworks vs Managed Agent Platforms, Securing Agent Frameworks Guide, Agent Identity and Access Guide, MCP Security Guide, Frontier AI Models Compared.