You need an agent that answers staff questions from the HR policies in SharePoint. You open a studio, wire up knowledge sources, and ship it a fortnight later.
That site already had an agent. Every SharePoint site does, scoped to exactly that content, on day one.
Microsoft offers four tiers, and most teams start further down than they need to.
Four tiers, from an agent that already exists to one you host yourself.
| Tier | Tool | Publishes to |
| Ready-made | Every site's built-in agent, the Agent Store, and the Word, Excel, and PowerPoint agents | Wherever that app already is |
| No-code | Agent Builder | Microsoft 365 Copilot, Teams |
| Low-code | Copilot Studio (Recommended) | Teams, web, WhatsApp, Slack, SMS, and more |
| Pro-code | Agents Toolkit | Microsoft 365 Copilot, Teams |
| Pro-code | Microsoft 365 Agents SDK | Any surface you code |
| Pro-code | Microsoft Foundry | Teams, Microsoft 365 Copilot, or any app via its endpoint |
Work down the tree and stop at the first thing that fits.
flowchart TD
Start(["You need an agent"]) --> Q0{"Does an agent that already<br>exists answer it?"}
Q0 -->|Yes| RM["Ready-made<br>Build nothing"]
Q0 -->|No| Q1{"Is it just for you or<br>a small team?"}
Q1 -->|Yes| AB["Agent Builder<br>No code"]
Q1 -->|No| CS["Copilot Studio<br>Low code, Recommended"]
CS --> Q2{"Does anything force<br>you past it?"}
Q2 -->|No| Ship(["Ship it"])
Q2 -->|"Source control and CI/CD"| TK["Agents Toolkit<br>Pro code"]
Q2 -->|"Lives inside a product you built"| SDK["Agents SDK<br>Pro code"]
Q2 -->|"Your own model or Azure networking"| FD["Microsoft Foundry<br>Pro code"]Most teams should stop at Copilot Studio. The common failure is not picking a weak tool, it is reaching for a powerful one before anyone has proven the agent gets used.
Note: A SharePoint agent is not a separate platform. It is a no-code agent authored inside SharePoint and scoped to that site, on the same orchestrator Agent Builder uses. Author it there when the content is one site and you want the site owner to keep it. Agent Builder covers the same SharePoint content plus OneDrive, Teams chats, and email.
Because it is a finished product rather than a toolkit. Knowledge from SharePoint, Dataverse, documents, and Microsoft Search connectors without code. Generative answers on by default. Per-user security trimming, so each person only sees content their Entra ID account can already open. File uploads, a long list of channels, and real dev, test, and production environments.
Reach and governance decide this, not features. Choose it when the agent serves a department, the organisation, or customers, or when it has to survive an audit.
It is the successor to Bot Framework, which Microsoft retired in December 2025, and it inherits that job: move messages between a user and whatever logic you wrote. Microsoft is blunt that it is not an AI model, not an orchestration engine, and not a no-code builder.
Picking it instead of Copilot Studio means building generative answers, grounding, and security trimming yourself. It ships a Copilot Studio client for .NET, JavaScript, and Python, so put Copilot Studio or Foundry behind it rather than in place of them.
Warning: Foundry is closer to Copilot Studio than the names suggest, using the same Microsoft 365 retrieval stack and the same permission trimming. Two catches: its SharePoint tool stops working once the agent is published to Teams, and every end user needs a Microsoft 365 Copilot licence.
Grounding and per-user permission trimming only work where your content and your identity provider already live, so the platform follows the data.
| Your content mostly lives in | Look at |
| Microsoft 365, SharePoint, Dataverse | Copilot Studio |
| Google Workspace or Google Cloud | Gemini Enterprise Agent Platform |
| AWS | Amazon Bedrock AgentCore |
| Salesforce | Agentforce |
| Nowhere in particular | A code-first SDK such as the OpenAI Agents SDK or LangGraph |
Note: Choose for where your data lives, not for the authoring experience. OpenAI shipped a visual Agent Builder in October 2025 and deprecated it eight months later, with shutdown set for 30 November 2026. Drag-and-drop canvases come and go. The grounding layer is what you are really committing to.