This guide is for founders, CMOs and business leaders ready to run AI agents of their own, and for the people who will build them. It is the last of five parts, and it describes the platforms as of October 2026.
What platforms can you use to run your own AI agents?
In 2026 there are three kinds of platform for running your own AI agents. Hosted agent services, such as Claude Managed Agents, Microsoft Foundry Agent Service, Google's Gemini Enterprise Agent Platform and Amazon Bedrock AgentCore, run the agent, its workspace, schedule and logs for you, and are the best start for most businesses. Frameworks such as the Claude Agent SDK, the OpenAI Agents SDK and Google's Agent Development Kit let developers run agents on your own infrastructure. No-code builders such as Copilot Studio, Zapier Agents and n8n let business teams build their own. You can then bring your own agent to Microsoft 365, Google Workspace, Slack, ChatGPT and your website through the open MCP and A2A standards.
AI Summary
Once you have an agent of your own, it needs somewhere to run. In 2026 there are three kinds of platform: hosted agent services, where a provider such as Anthropic, Microsoft, Google or AWS runs the agent for you; frameworks your developers run on your own servers; and no-code builders for teams that do not write code. For most businesses, a hosted service is the best start. Then bring your own agent to where people already work: Microsoft 365, Google Workspace, Slack, ChatGPT and your own website now accept outside agents through two open standards, MCP and A2A. Choose a platform for its fit with your data and team, keep your agent's job description, instructions and tests portable, and govern every agent like an employee with an identity, permissions and a budget.
Background
Part 1 defined an agent. Part 2 covered how to build one, Part 3 where it should run, and Part 4 what you can buy. This final part is about agents you own: the platforms that run them, and how to put them where your people and buyers already are.
What an Agent Platform Does
An agent needs more than a model. A platform supplies the rest:
- The loop. Running the agent step by step, calling tools, handling errors and knowing when to stop.
- A workspace. A safe, private place to read files, run code and browse, separate from your other systems.
- Connections. Access to your tools, usually through MCP connectors, and credentials stored safely.
- Memory. What the agent keeps between runs.
- Schedules and triggers. Running every morning, or whenever a lead arrives.
- Identity and permissions. Who the agent is, what it may touch, and on whose behalf it acts.
- Logs and evaluation. A record of every step, and tools to test changes against your real cases.
You can build all of this yourself, and some companies should. Most should rent it.
Three Kinds of Platform
The Major Platforms
Anthropic
Claude Managed Agents, in public beta since April 2026, hosts your agent and a private workspace for it, runs it on a schedule, and stores credentials in vaults so secrets never enter the agent's workspace. You define the agent once and start sessions as work arrives. The Claude Agent SDK gives developers the engine behind Claude Code to run on their own infrastructure. Claude Cowork and Claude Code are where people work beside agents day to day. Anthropic also created MCP, the standard for connecting agents to tools, which it donated to the Linux Foundation in December 2025.
OpenAI
The Agents SDK and the Responses API are OpenAI's foundation for agents you build. ChatGPT Workspace Agents, in research preview, let teams build agents inside ChatGPT that run on schedules. A caution worth learning from: OpenAI deprecated its visual Agent Builder in June 2026, a year after launch, with shutdown set for November 30. Teams that built there must move. Choose platforms knowing that some will change under you.
Since April 2026, Vertex AI has become the Gemini Enterprise Agent Platform, with a low-code Agent Studio, the open-source Agent Development Kit (ADK), a hosted runtime, memory, agent identity and a registry. Google created the A2A protocol, now a Linux Foundation project at version 1.0, and Gemini Enterprise accepts partner agents over A2A.
AWS
Amazon Bedrock AgentCore, generally available since October 2025, provides the building blocks as separate services: a runtime, a gateway for tools, memory, identity, a browser, a code interpreter, policy, evaluations and a registry. It works with agents built on any framework and model, and in August 2026 added a way for agents to pay for APIs and content.
Microsoft
Microsoft Foundry, formerly Azure AI Foundry, runs hosted agents for developers, and Copilot Studio lets business teams build agents for Microsoft 365. Agent 365, generally available since May 2026, is a control plane to register, secure and govern agents from any vendor, including ones built on AWS and Google.
No-code and low-code
Copilot Studio, Zapier Agents, n8n, Make, Lindy and Relevance AI let teams assemble agents from connectors and plain-English instructions. They are a fast way for marketing and operations teams to automate their own work. Watch for two things as an agent grows: costs, which often scale with every action, and the point where a job outgrows what a visual builder can test and govern.
Bring Your Own Agent
The biggest change of 2026 is that the places people work now accept outside agents. You no longer have to choose between building on Microsoft, Google or Salesforce. You can build an agent once and bring it to each of them:
Two open standards make this possible:
- MCP, the Model Context Protocol, connects an agent to tools and data. It began at Anthropic and is now governed by the Linux Foundation's Agentic AI Foundation. Nearly every major platform supports it.
- A2A, the Agent2Agent protocol, lets agents from different companies talk to each other. It began at Google, is now a Linux Foundation project at version 1.0, and is supported by more than 150 organizations, including Microsoft and AWS.
Standards for agent payments and agent shopping are earlier and still shifting. Google gave its Agent Payments Protocol to the FIDO Alliance in 2026, and OpenAI ended checkout inside ChatGPT in March 2026 in favor of sending buyers to retailers' own sites. Watch them, but do not build a strategy on them yet.
Your website is one of the places you can bring an agent. Our own A2A agent is listed in an agent card on buildmarketing.ai, so any A2A-compatible agent can find it and ask it questions. You can see it working.
How to Choose a Platform
- Start where your data is. If you run on Microsoft 365, Google Workspace or AWS, that provider's platform is the shortest path. If you do not, a model provider's hosted service, such as Claude Managed Agents, keeps you independent.
- Match the team. No-code builders for business teams. Hosted services for a small technical team. Frameworks for engineers building agents into a product.
- Check the guardrails. Credentials kept out of the agent's reach, permissions you can narrow, approval steps, full logs, spending caps and the ability to run your test cases.
- Insist on the standards. MCP for tools and A2A for other agents keep your options open.
- Plan to move. Keep the job description, instructions, knowledge and test cases in your own files. They are the agent; the platform is where it runs today.
Govern Agents Like Employees
As agents multiply, treat each like a member of staff. Give it an identity, so every action is traceable to it. Give it only the permissions its job needs. Give it a budget, with alerts. Give it an owner who reviews its work. Keep a register of every agent you run, bought or built, and what it can touch. Platforms such as Agent 365, the Gemini Enterprise Agent Platform's registry and Bedrock AgentCore now provide much of this, and it is worth turning on from your first agent.
What We Use
Our site's agents are small serverless functions on Netlify that call Claude through Anthropic's API, with their knowledge drawn from our own pages and their skills switched on in our admin. That is the build-your-own path at its simplest, and right for a small business with one developer's worth of agents. If we needed long-running agents with workspaces and schedules, a hosted service such as Claude Managed Agents would be the next step. The people side runs on Claude Cowork and Claude Code, which is also how this series was produced, with every change reviewed before it went live.
Test yourself
5 quick questions. Pick an answer to see if you are right.
1. What is the best first platform for most businesses running an agent of their own?
Show answer
The answer is A hosted agent service. Hosted services from Anthropic, Microsoft, Google or AWS run the loop, workspace, schedule and logs for you, so you focus on the agent itself.
2. What does it mean to bring your own agent?
Show answer
The answer is Building an agent once and connecting it to places like Microsoft 365, Slack, ChatGPT or your website. Many platforms now accept outside agents through open standards, so one agent can work wherever your people and buyers are.
3. Which two open standards make bringing your own agent possible?
Show answer
The answer is MCP and A2A. MCP connects agents to tools and data. A2A lets agents from different companies talk to each other.
4. OpenAI deprecated its visual Agent Builder a year after launch. What is the lesson?
Show answer
The answer is Keep your job description, instructions and tests portable. Platforms change. If your agent's definition lives in your own files, moving platforms is a migration, not a rebuild.
5. How should you govern a growing number of agents?
Show answer
The answer is Treat each like an employee, with an identity, narrow permissions, a budget and an owner. An identity, least access, a budget, an owner and a register of every agent keep you in control as agents multiply.
Now put it to work on your own website. Run MachineReady to see how AI agents read it, or send our A2A agent to see if it can talk to theirs. Both are free.
Conclusion
Which platforms should you use to run your own AI agents? For most businesses, a hosted agent service from Anthropic, Microsoft, Google or AWS, with no-code builders for teams that automate their own work and frameworks for agents inside your product. Then bring your agents to where people already work, through MCP and A2A. Keep your agent's job description, instructions and tests portable, and govern every agent like an employee.
That completes the series. Start with Part 1 if you are sharing it with your team, and take the quizzes together.
Is your website ready for your own agent? Send our A2A agent to your site to see if it already speaks A2A, or run MachineReady for a full report. Get in touch when you are ready to run an agent of your own.
← Back to all articles