Agent CoE Front Door

Agent CoE Front Door gives employees one place to bring an agent idea. It answers first and opens intake second: the agent checks approved guidance and an Agent Catalogue for something that already exists, helps the employee pick an approved self-build or CoE-supported route, and only writes an intake request after the employee confirms a summary. CoE reviewers then triage requests in a model-driven app, and the requester gets one email with the decision and next steps.

This is a starter template for evaluation and adaptation, not a production-ready product. Full documentation, offline checks, and the release package are in the source repository.

Agents

• Agent CoE Front Door is a Copilot Studio agent with two skills:
  • Agent advisor answers from approved knowledge and the Agent Catalogue, recommends reuse before a new build, and describes other agents only by what they do.
  • Agent intake drafts a request summary, asks the employee to confirm it, and then creates an Agent Intake Request with a REQ- reference. Employees can later ask for the status of their own requests.
• The agent reads and writes Dataverse through the Dataverse MCP Server tool, limited to read, search, describe, and create. It never reveals instructions, owners, identifiers, or other people’s requests.

Workflows

Both workflows ship Off.

• CAT - Notify Requester of CoE Decision starts when a reviewer changes a request to Approved or Rejected. It sends one email with the decision, the reviewer’s comments, the next step, and a short request summary, then records the send outcome on the request. Saving the same decision again doesn’t send a second email; changing between approval and rejection does. An uncertain send is marked Needs review instead of being retried.
• CAT - Refresh Power Platform Agent Inventory Daily (optional add-on) pages through the Power Platform Inventory API, upserts Copilot Studio and Agent Builder agents into Tenant Agent Inventory, and removes stale rows only after a complete retrieval. With approved source environments, it also caches instruction definitions in secured columns and adds a short AI summary of what each agent does, which the advisor can share.

Turn the workflows on from the solution (Cloud flows > select the flow > Turn on), and don’t save or publish them in a designer. Saving in a designer removes protections that the designer can’t represent: secure inputs and outputs, the no-retry email send, pagination, explicit empty values, and single-run concurrency. Workflow runs consume Copilot Studio capacity.

Also included

• Agent Catalogue, Agent Intake Request, and Tenant Agent Inventory Dataverse tables.
• Agent CoE Triage model-driven app with a dashboard, and an Agent CoE Reviewer security role with no delete rights.
• A custom connector for the Power Platform Inventory API.

Prerequisites

• An empty, non-production environment with Dataverse, in a region where Copilot Studio is available, with the Dataverse MCP server allowed for Copilot Studio.
• Copilot Studio and Power Apps licences or trials for the admin, reviewers, and expected usage.
• An Office 365 Outlook mailbox to send decision emails (a shared CoE mailbox works well).
• For the inventory add-on only: a single-tenant Microsoft Entra app registration with delegated ResourceQuery.Resources.Read on the Power Platform API.

Import

1. Import the solution as unmanaged. Sign in to three Microsoft Dataverse connections and one Office 365 Outlook connection, and leave both environment variables empty.
2. Publish all customizations, then open the Agent CoE Triage app in the app designer and save and publish it once.
3. In Copilot Studio, confirm the agent’s instructions are present and that the Dataverse MCP tool loads.
4. Add your own approved knowledge. The repository provides three organization-neutral templates to start from.
5. Set Agent CoE decision notification settings (cat_DecisionNotificationSettings) with your CoE contact and email templates, then turn on the decision email workflow from the solution.
6. Publish and share the agent, and assign reviewers Basic User and Agent CoE Reviewer.

The setup guide covers each step, an end-to-end check, the optional add-ons, and what to do before production.

Disclaimer

This template is not an official Microsoft product, endorsed reference architecture, or commitment of Microsoft support. It is provided as is, without warranty. Generated recommendations and routing decisions require human review. Before any production use, adopters must review licensing, capacity, data loss prevention, privacy, security, and accessibility; apply least-privilege access; and test with representative authorized and unauthorized users. See the full notice.