Where Microsoft is going with agents

Microsoft’s strategy is enterprise-first agents in the M365 and Azure estate. That isn’t a single product — it’s a spectrum:

• Copilot Studio (low-code agent builder; runs on the Power Platform).
• Foundry / Azure AI Foundry (full-code, multi-model, evaluation-led; the platform side).
• Connectors + Declarative Agents (pre-built integrations into M365).
• Agent 365 (enterprise governance plane — usage, audit, identity for agents themselves; GA imminent).
• Apps SDK plugins for ChatGPT (Microsoft is investing here too — distribution where users are).

Every one of these has converged on MCP as the tool-connection layer. Foundry agents speak MCP. Copilot Studio added MCP support. VS Code’s GitHub Copilot uses MCP. That convergence is genuine; it’s the most important architectural decision Microsoft’s made in this space.

Where they lead

• Enterprise IT integration. If your data lives in M365 / SharePoint / Dataverse, agents that touch it via Microsoft’s connectors avoid a long tail of bespoke plumbing.
• Governance + identity. Entra-based identity for agents themselves (Agent 365) is a real differentiator — you can audit which agent did what, attach RBAC, rotate credentials, scope permissions. No other vendor matches this depth on the management plane.
• Multi-product reach. Same agent author can publish to Teams, Outlook, M365 Chat, ChatGPT, even external surfaces.
• Hybrid-cloud paths. For regulated industries that need on-prem / sovereign cloud, Microsoft has clearer paths than Anthropic / OpenAI.

Where they lag

• Pricing complexity. Multiple SKUs, tier-gated features, and add-ons make total cost-of-ownership genuinely hard to model up front. Get a Microsoft FastTrack engagement before sizing a deployment — the math is non-obvious.
• Brand / pace lag on protocols. Anthropic ships protocol updates faster. By the time Microsoft adopts a new MCP spec version, Claude Desktop and Cursor have shipped support already. This is fine — Microsoft optimises for stability — but it’s a real lag if your engineering team is moving at startup pace.
• Inside-the-fence default. Microsoft’s defaults assume your data is in M365 / Azure. Agents that need to reach across into AWS / GCP / arbitrary SaaS are workable but not first-class.

Honest take

Microsoft’s the right default if your org already runs on M365 / Azure and you need the governance story. Agent 365’s audit + identity model is genuinely ahead of every other vendor on the management plane. The product depth covers everything from low-code business builders to high-code Foundry workflows.

If you’re a small team building from scratch with no Microsoft footprint, the licensing model is heavier than you need. Anthropic + MCP is lighter to start, and you can add Microsoft surfaces later via Apps SDK or connectors when the business case appears.

Further reading

• Microsoft Copilot Studio — microsoft.com/copilot-studio
• Azure AI Foundry — azure.microsoft.com/foundry
• Agent 365 — microsoft.com/agent-365 (search for Agent 365)
• MCP support roadmap — github.com/microsoft (search for “mcp”)
• Sush’s MS-side blog posts — aguidetocloud.com (Microsoft Copilot tag)

Last reviewed 2026-05-07. Next quarterly review due 2026-08-07. Disclosed: Sush works at Microsoft NZ — independent take.