The Power Platform Center of Excellence for AI agents looks fundamentally different in 2026 than it did even twelve months ago - and most enterprise governance programs have not caught up. In May 2026, Microsoft confirmed that the long-standing CoE Starter Kit is no longer actively maintained, with its core capabilities now delivered natively in the Power Platform admin center through Inventory, Usage, Monitor, and Actions experiences. Issues are no longer being reviewed or addressed.

For enterprise architects and platform leads, this is not a minor tooling update. It is a structural shift in how a Power Platform Center of Excellence for AI agents should be designed, staffed, and operated - particularly as Copilot Studio agents become the fastest-growing workload class on the platform.

What Is a Power Platform Center of Excellence for AI Agents?

A Power Platform Center of Excellence (CoE) for AI agents is the governance, enablement, and operational function that establishes how an enterprise builds, deploys, and operates AI agents at scale on Microsoft Copilot Studio and the broader Power Platform. In 2026, it is built on native Power Platform admin center capabilities - Inventory, Usage, Monitor, and Actions - combined with Microsoft Entra Agent ID, managed environments, and Advanced Connector Policies, rather than on the legacy CoE Starter Kit.

Table of Contents

• The 2026 Shift: From Starter Kit to Admin Center
• The Four Pillars of a Modern CoE
• What's Different When Agents Are Involved
• The Operating Model: Roles and Responsibilities
• A 90-Day Implementation Plan
• Frequently Asked Questions

The 2026 Shift: From Starter Kit to Admin Center

For nearly six years, the CoE Starter Kit was the de facto governance toolkit for Power Platform. Most existing enterprise CoE designs reference it directly. That foundation is now eroding underneath them. Per the official transition guidance, the CoE Starter Kit remains available for existing and new deployments but will not receive new capabilities, and issues are no longer being addressed.

The replacement is not a single tool - it is four native admin center experiences:

• Inventory - View and govern all apps, flows, and agents created across the tenant.
• Usage - Track adoption and identify top resources and their owners.
• Monitor - Track the operational health of heavily used resources.
• Actions - Identify risks, enforce best practices, and take action on governance insights across the tenant.

In addition to the UI, capabilities are available through the Microsoft Power Platform CLI, the Power Platform API, the Power Platform inventory API, and the Power Platform for Admins V2 connector - which together enable programmatic CoE workflows that were previously possible only through custom Starter Kit flows.

Why the Change Matters for AI Agents

The Starter Kit was designed for an era of canvas apps and cloud flows. It was retrofitted for Copilot Studio agents but never natively designed for them. The native admin center experiences, by contrast, treat agents as first-class resources alongside apps and flows. That alignment is what makes a modern Power Platform Center of Excellence for AI agents operationally tractable. Agents appear in the same inventory, governance, and action workflows as everything else makers build - without the integration debt of the Starter Kit.

The Four Pillars of a Modern CoE

The structural change does not eliminate the historical CoE disciplines. It re-platforms them. A modern Power Platform CoE for AI agents rests on four pillars, each of which now maps directly to native admin center capabilities.

1. Visibility

Visibility is the prerequisite for everything else. You cannot govern what you cannot see. The Inventory experience provides a tenant-wide view of every app, flow, and agent - including ownership, environment, and connector usage. For agents specifically, this is the first time enterprise platform teams have had a single pane of glass for agent sprawl detection without building custom telemetry.

2. Adoption Intelligence

Adoption metrics tell you which agents are creating real value and which are quietly idle. The Usage experience surfaces top resources by activity and their owners, enabling CoE teams to identify both standout success stories worth amplifying and dormant agents worth retiring. For finance leaders trying to justify platform investment, this is the data layer that converts platform spend into business outcomes.

3. Operational Health

Agents fail differently than apps. They drift, hallucinate, and degrade with model updates. The Monitor experience tracks operational health of heavily used resources, allowing the CoE to spot reliability issues before they become user-facing incidents. Integrated with Application Insights and Power Platform telemetry, Monitor becomes the source of truth for agent SLO definition.

4. Enforcement

Insight without enforcement is theater. The Actions experience converts governance signals into action - revoking access, applying DLP policies, archiving unused resources, and surfacing risks for review. For AI agents, this is where governance moves from advisory to operational.

The shift from CoE Starter Kit to the Power Platform admin center is less about new tooling and more about who owns governance. When governance is native, it stops being a side project run by a single architect and becomes operational telemetry the whole platform team consumes.

What's Different When Agents Are Involved

Apps and flows have predictable behavior. Agents do not. A CoE designed only for traditional Power Platform workloads will fail to govern agents effectively for five specific reasons.

1. Identity Is Now an Entra Concern

Every Copilot Studio agent now receives a unique Microsoft Entra Agent ID, which means agent identity governance lives in Entra - not in Power Platform admin center alone. A modern CoE must operate across both surfaces. This is a meaningful change for CoE staffing: identity engineers belong in the CoE forum, not just Power Platform admins.

2. Connector Permissions Are Agent-Specific

When an agent is published, its connector permissions attach to its Entra Agent ID as first-class API permissions. The CoE must define which connectors are approved for agent use, and which require additional Conditional Access controls. This is a new governance artifact that did not exist for traditional apps.

3. MCP Servers Expand the Attack Surface

Model Context Protocol servers let agents reach external tools and data. They are easy to create and easy to over-permission. The CoE must maintain an MCP allow-list, treat MCP onboarding as a high-privilege review, and integrate MCP telemetry into the Monitor experience. Most CoEs have not added this discipline yet.

4. Evaluation Replaces Testing

Apps are tested. Agents are evaluated. Forrester data shows that 64% of leaders cite evaluation gaps as the top blocker to agent production. The CoE must establish a shared evaluation framework - covering grounding quality, response accuracy, task completion, and safety - that every production agent passes before publish.

5. Lifecycle Is Faster and More Volatile

Agents are spun up, modified, and retired faster than apps. CoE processes that worked on a quarterly review cadence will not keep up. The CoE needs continuous, automated lifecycle controls - most of which are now expressible through Actions and the Power Platform API.

The Operating Model: Roles and Responsibilities

A Power Platform CoE for AI agents is no longer a single team. It is a federated function with clear ownership across four roles. At BabyBots, we consistently see this four-role pattern emerge in successful enterprise engagements, regardless of organization size.

The Platform Owner

Owns the Power Platform admin center, environment strategy, and DLP policies. Accountable for Inventory, Usage, Monitor, and Actions configuration. Typically reports into Enterprise Architecture or Modern Workplace.

The Identity Owner

Owns Entra Agent ID enablement, blueprint governance, and Conditional Access policy design for agents. Accountable for integrating agent identities into existing access reviews. Reports into Identity and Access Management.

The Agent Owner (per agent or agent family)

Owns the business purpose, evaluation criteria, and lifecycle of individual agents. Forrester data shows that 56% of enterprises now name a dedicated agent owner in 2026, up from 11% in 2024 - and ownership maturity correlates strongly with production success.

The Security Operations Lead

Owns agent telemetry integration into XDR, DLP enforcement, MCP allow-listing, and prompt injection detection. Accountable for treating agent incidents with the same operational rigor as user account compromises.

A 90-Day Implementation Plan

The fastest path to a functioning CoE for AI agents is sequential, not parallel. Treating governance, identity, and enablement as concurrent workstreams typically results in misaligned policies and rework.

Days 1–30: Establish the Baseline

• Stand up the Power Platform admin center experiences. Enable Inventory, Usage, Monitor, and Actions across all production environments.
• Inventory all existing agents. Identify ownership, business purpose, and connector usage. Flag agents without named owners.
• Convene the CoE forum. Platform owner, identity owner, security ops lead, and a rotating agent owner representative.
• Define managed environment strategy. Decide which environments support agent development versus production.

Days 31–60: Identity and Policy

• Enable Entra Agent ID in non-production environments. Validate audit log integration and connector permission visibility.
• Publish the connector allow-list. Define which connectors are approved for agent use and which require additional review.
• Build the agent evaluation framework. Specify the gates every production agent must pass - grounding, accuracy, completion, safety.
• Define DLP and Advanced Connector Policies specific to agent workloads.

Days 61–90: Operationalize and Scale

• Roll out Entra Agent ID to production. Add agent identities to existing access review cadences.
• Integrate agent telemetry into the SOC. Feed Monitor signals into XDR for correlated detection.
• Publish the agent maker enablement program. Documentation, training, and review checkpoints for citizen developers building agents.
• Establish quarterly CoE health reviews using Usage and Monitor data as the primary inputs.

Key Takeaways

• The CoE Starter Kit is no longer actively maintained. Modern CoEs are built on native Power Platform admin center capabilities.
• Four experiences - Inventory, Usage, Monitor, and Actions - replace the legacy Starter Kit governance functions.
• AI agents introduce five governance differences that traditional CoE designs do not address: Entra-based identity, connector permissions, MCP risk, evaluation rigor, and lifecycle volatility.
• A modern CoE is a federated function with four roles: platform owner, identity owner, agent owner, and security operations lead.
• A 90-day implementation plan moves sequentially from baseline visibility, to identity and policy, to operational scale.

Frequently Asked Questions

Is the Power Platform CoE Starter Kit still supported?

The CoE Starter Kit remains available for existing and new deployments but is no longer actively maintained. It will not receive new capabilities, and issues are no longer reviewed or addressed. Microsoft is investing in native Power Platform admin center experiences - Inventory, Usage, Monitor, and Actions - as the replacement.

What replaces the CoE Starter Kit?

The Power Platform admin center now delivers core CoE capabilities natively. Inventory provides tenant-wide visibility, Usage tracks adoption, Monitor surfaces operational health, and Actions enforces governance. The Power Platform CLI, Power Platform API, and Power Platform for Admins V2 connector provide programmatic access.

How is a CoE for AI agents different from a traditional Power Platform CoE?

A CoE for AI agents must extend across both Power Platform admin center and Microsoft Entra ID, govern connector permissions on agent identities, maintain an MCP server allow-list, run continuous agent evaluation rather than one-time testing, and operate on a faster lifecycle cadence than traditional apps and flows.

Who should sit in a modern Power Platform CoE forum?

The four core roles are platform owner (admin center and environment strategy), identity owner (Entra Agent ID and Conditional Access), agent owner (business purpose and evaluation for specific agents), and security operations lead (XDR, DLP, and MCP allow-listing).

How long does it take to stand up a CoE for AI agents?

A working baseline can be established in 90 days using a sequenced plan: baseline visibility in the first 30 days, identity and policy in days 31–60, and operational scale in days 61–90. Full maturity - including continuous evaluation and integrated security operations - typically takes two to three quarters.

The Strategic Question

The Power Platform CoE was historically a governance function. In 2026, with agents scaling faster than human users and identity, security, and platform decisions converging on a single workload class, it is becoming the operating model for enterprise AI itself. Will your CoE evolve into that operating model, or remain a Power Platform side project as the agent population overtakes the rest of your platform combined?