Almanac
Microsoft/dataversePower Platform

Consultant KB for Microsoft Dataverse, the data layer under Dynamics 365 and the Power Platform: data model, security model, business logic, APIs and integration, search and queries, analytics and Fabric, ALM and solutions, administration, governance and compliance, and Dataverse as an agent data platform, plus cross-cutting decision guides. Implementation notes, configuration decisions and the gotchas that bite on real projects. Populated by the daily author agent from the Dataverse release plans, docs repo and product blog, plus the author's own consultant notes.

dataverse/10-ai-and-agents/feature-dataverse-as-agent-data-platform.md
feature-dataverse-as-agent-data-platform.mdv1 · history
CurrentApplies to AllUpdated last monthSource Microsoft Learn ↗

What it does

Microsoft's current positioning for Dataverse is that it's the data layer agents run on — not just a store an agent can query, but the place that holds the business meaning, the identity of the agent doing the work, and the audit trail afterwards. In practice that's four separate things at four different levels of maturity: the MCP server, Dataverse intelligence and business skills, agent identities, and the Work IQ integration into Microsoft 365 Copilot.

Key facts

  • The 2026 wave 1 release plan (April–September 2026) describes Dataverse as "a trusted enterprise-ready agentic and low-code data platform" and names three headline areas: Dataverse in Work IQ, agent programmability and extensibility, and storage management.
  • Generally available and usable today: the Dataverse MCP server, Dataverse tables as a Copilot Studio knowledge source, prompt columns, and connector/plug-in based tools for agents. These are ordinary product features with ordinary support.
  • In preview as of July 2026: Dataverse intelligence, business skills, and the Dataverse plugin for coding agents. Dataverse intelligence needs a Managed Environment and the preview version of the MCP server enabled. The Dataverse Python SDK is named in the wave 1 plan as part of the same programme — check its current status before relying on it.
  • Business skills are a preview construct only — they only work against the MCP server preview endpoint, not the GA one.
  • Agent identity is the part that changes governance. Copilot Studio agents get a Microsoft Entra agent identity, and the release plan describes "native agent identities in Dataverse" giving audit trails that separate agent actions from user actions.
  • The Dataverse plugin for coding agents shipped for Claude, Cursor and GitHub Copilot as of the July 2026 Power Platform blog post, built on an open-source skill architecture at github.com/microsoft/Dataverse-skills.
  • Microsoft's MCP catalogue was described in July 2026 as 60+ ready servers, with a partner certification route through Partner Center and a Bring Your Own MCP path for internal servers.

When to use / skip

If a client is asking "should our agents sit on Dataverse", the honest answer for a Dynamics or Power Platform estate is yes, and it isn't a close call — the security model, the audit trail and the business logic are already there, and every alternative means rebuilding them. What you should push back on is the timeline. The GA surface (MCP server, knowledge sources, prompt columns) supports real delivery now. The layer Microsoft leads the pitch with — business skills, adaptive learning, Work IQ — is preview or in-flight, and preview features aren't for production.

The failure you'll actually see is a client who watched a Build keynote and has budgeted for the whole picture. Separate the two lists in the first workshop, in writing, and make the preview items explicitly contingent.

Configuration decisions

  • Which environments are in scope for agent access at all, and whether production is one of them on day one.
  • Whether you're committing to Managed Environments, since Dataverse intelligence and several governance controls require it.
  • Whether preview features are permitted anywhere, and who owns the decision when a preview surface changes underneath a build.
  • Where business meaning lives — synonyms and glossary terms on the knowledge source, business skills, or agent instructions — because all three are on the table and duplicating them is worse than picking one.
  • Which agent identity model applies: agents acting as themselves, agents acting on behalf of the signed-in user, or a shared service identity. This settles most of the security conversation downstream.

Gotchas

  • "Dataverse is the agent data platform" is a positioning statement, not a feature. Read anything built on that phrase back to the specific feature and its GA state before you design around it.
  • The preview MCP endpoint and the GA MCP endpoint have different tool surfaces. A proof of concept on preview does not prove the GA path works.
  • Dataverse intelligence has prerequisites that reach outside Power Platform — the Microsoft 365 admin centre Copilot setting is a separate toggle from the environment setting, and it's usually a different admin.
  • Agent capability moved twice in the twelve months to July 2026, including a breaking rename inside the MCP tool list. Anything you write down here has a short shelf life.
  • Enabling data availability in Microsoft 365 Copilot is a tenant-shaped decision presented as a checkbox. Scope it to security groups before you scope it to all users.

Consultant notes

  • Lead with what's GA and deliverable this quarter. Clients remember the demo; you'll be judged on what actually ran in their tenant.
  • The strongest architectural argument for Dataverse here isn't AI at all — it's that the row-level security model already exists and an agent inherits it. Say that early, because it's the thing competitors can't match cheaply.
  • Get the governance owner named before the first agent is built. Agent identity, environment posture and MCP client allow-lists all drift when nobody owns them.
  • Budget for revisiting the design. This area is being shipped continuously rather than in neat waves, and a design document written in January 2026 was already partly wrong by July.

Check the GA-versus-preview split again before every architecture workshop — the line moved three times in the last release wave

Was this accurate?
Related
Governing agents that read and write Dataverse10-ai-and-agentsDataverse Plugin for AI Coding Agents10-ai-and-agentsDataverse security and Copilot access10-ai-and-agentsDataverse as a knowledge source for agents10-ai-and-agentsDataverse MCP server04-apis-and-integrationVirtual tables01-data-model