Govern agent lifecycle and costs
Your work doesn't end when you publish an agent. You still need to control updates, access, data use, and costs.
This unit introduces governance and application lifecycle management (ALM). You learn why these practices matter and when to ask an administrator for help. You don't build a deployment pipeline in this module.
Understand application lifecycle management
Application lifecycle management covers how you build, test, release, and update an agent. The main concepts are:
Solutions
A solution groups an agent with related items, such as workflows and connections. You can move the solution from a development environment to a test or production environment.
Use a solution from the start when other teams depend on the agent or when releases need formal review.
Environment strategy
Many organizations use separate environments for development, testing, and production. This approach keeps unfinished changes away from live users.
You usually build in a development environment. An administrator or pipeline moves an approved version to test and production. Settings and connections might need different values in each environment.
Environment variables and connection references
When an agent moves between environments, configuration values and connection details often differ. A SharePoint site URL in development might not be the same as in production.
- Environment variables store values that differ between environments, such as a site address.
- Connection references let each environment use the correct connector account.
Set up these items early when an agent connects to other services. Adding them after release takes more work and can cause errors.
Pipelines and deployment automation
Power Platform Pipelines move solutions between environments. A pipeline can also require approval before a release.
As a maker, you might submit a deployment request and test the agent after deployment. An administrator or IT team usually configures the pipeline.
Data loss prevention policies
Data Loss Prevention (DLP) policies control which connectors agents can use and which data can flow between services. An administrator configures DLP policies at the environment or tenant level. If your agent uses a connector that a DLP policy restricts, you see an error during authoring or at runtime. Contact your administrator to request an exception or to identify a compliant alternative connector.
Role-based access and sharing
Role-based access control determines who can view, edit, share, and interact with agents. Makers control sharing at the agent level through Copilot Studio's sharing settings. Administrators enforce broader access policies through environment roles and security groups.
Auditing records supported Copilot Studio activities, including creation, modification, sharing, and deletion events. Administrators with the required permissions search these activity logs in Microsoft Purview > Audit. Auditing requires the appropriate licensing, configuration, and roles. Use the Power Platform admin center for environment, capacity, and consumption administration, not as the destination for searching these audit events. For details, see Power Platform activity logs and auditing in Microsoft Purview.
Before you share an agent broadly, confirm:
- Authentication and connection settings use the intended identity, with only approved access to knowledge sources and tools. For user-authenticated connections, intended users need permission to the underlying resources.
- Whether your organization requires administrator approval before the agent is available in Teams or Microsoft Copilot, and which approval process applies.
- Regulatory or compliance requirements don't limit the data the agent can access or the users it can serve.
Understand Copilot Credits
Copilot Credits measure consumption in Copilot Studio. For agents powered by the Microsoft Copilot harness, charges cover language model tokens, tools (including knowledge and MCP tools), and the harness itself. Billing can begin during building, before publication. An administrator manages prepaid capacity, pay-as-you-go billing, and consumption controls for the environment.
What consumes Copilot Credits
Activities that can consume Copilot Credits include:
- Creating an agent or workflow through natural-language building.
- Testing the agent on the Preview tab.
- A user conversation where the agent calls its language model to generate a response.
- Knowledge retrieval and tool use, including MCP tools and workflow execution.
- Generating evaluation conversations.
- Running an evaluation, including generating agent responses and scoring them.
Tool charges aren't limited to model-assisted input interpretation or output generation. The amount consumed depends on the activity, model, token usage, and complexity of the operation. Microsoft publishes current consumption information in Overview of usage-based billing. Avoid relying on rates documented in training materials because pricing can change.
Distinguish manual edits from billable building and testing
Typing instructions or saving configuration isn't, by itself, a request to run the agent's model. However, don't treat all authoring as free or assume an unpublished agent can't consume credits. Natural-language building, Preview testing, and evaluation generation and execution can consume credits before you publish.
Note
Plan evaluation frequency with credit consumption in mind. Run evaluations at key milestones—before initial release and after significant changes—rather than after every minor edit. Use the Preview tab for quick checks during development.
Monitor capacity and consumption
Administrators monitor Copilot Credit consumption and available capacity in the Power Platform admin center. If your organization's credit balance runs low, agents may fail to respond or return capacity-related errors. If you see unexpected failures on the Monitor tab that don't trace to a specific tool or connection error, contact your administrator to check whether the issue is capacity-related.
Consumption can be monitored per agent as well as across environments and the tenant. Administrators can configure an agent-level monthly limit, notifications, and a Stop usage option. Environment-level allocations and pay-as-you-go settings determine the other capacity sources available. Deleting an agent doesn't refund or erase credits already consumed.
To limit unnecessary credit use, keep each test set focused on common scenarios. Run evaluations before release and after important changes. Use the Preview tab for quick checks while you build.
For more information, see Manage costs for agents powered by the Microsoft Copilot harness.