3.4 Agent Creation and Lifecycle

Quick Answer: Build custom agents in Microsoft Copilot Studio using topics, knowledge sources, and actions. Before org-wide release, agents pass through an admin approval process in the Microsoft 365 admin center. Monitor and govern the full lifecycle through the M365 admin center and Power Platform admin center.

Custom Agents vs Built-in Copilot

Microsoft 365 Copilot ships with the tenant — a general productivity assistant across M365 apps grounded on the semantic index. Custom agents are purpose-built assistants for specific business scenarios: IT helpdesk triage, HR policy Q&A, sales playbook guidance, or facility requests. Makers author them in Microsoft Copilot Studio, a low-code platform that requires no deep coding expertise for standard scenarios.

AB-900 tests the creation toolchain, governance gates, and lifecycle consoles — not advanced bot scripting.

Microsoft Copilot Studio Core Concepts

Topics

A topic is a modular conversation flow defining how an agent responds to a specific user intent. Each topic contains:

• Trigger phrases — utterances that activate the topic ("reset my password", "request PTO")
• Nodes — sequential steps: messages, questions, conditions, and actions

Topics are the building blocks of agent logic. Complex agents compose many topics rather than one monolithic script.

Knowledge Sources

Agents ground responses on configured knowledge sources:

Admins must verify sensitivity of connected sources because agents inherit the access patterns of configured knowledge.

Actions

Actions extend agents beyond conversation — connecting to external systems through:

• REST API calls
• Power Automate flow triggers
• Pre-built connectors (ServiceNow, Salesforce, etc.)

An action lets an agent retrieve CRM data, submit a ticket, or update a record — not merely chat about it.

Generative Answers

Generative answers let an agent query configured knowledge sources with generative AI rather than following only rigid topic flows — blending authored logic with retrieval-augmented responses.

Agent Types: Declarative vs Autonomous

Both use the same underlying model infrastructure; the distinction is architectural reasoning capability, not a specific GPT version or deployment channel.

Deployment Channels

Copilot Studio publishes agents to multiple channels:

The Teams channel makes a custom agent available directly as a Teams app — a frequent exam answer when the stem specifies Teams deployment.

Power Platform Runtime and Governance

Power Platform environments provide:

• Runtime infrastructure for Copilot Studio agents
• Dataverse for data storage and structured records
• Application lifecycle management (ALM) for dev/test/prod promotion
• DLP policies for connectors restricting which data sources agents can reach

The Power Platform admin center manages environments, maker permissions, tenant-level settings, and Copilot Studio governance. It complements — not replaces — the Microsoft 365 admin center.

Agent Approval Process

Before custom agents reach all users, organizations enable an agent approval process:

1. Maker builds and submits agent in Copilot Studio
2. Administrator reviews in Microsoft 365 admin center
3. Approve or reject based on security, data access, and compliance readiness
4. Published agents appear in the org catalog for authorized users

This prevents ungoverned agents from accessing organizational data or performing actions without review. Retention labels and conditional access policies do not substitute for this workflow.

Agent Lifecycle Management

End-to-end agent lifecycle management covers:

Create → Review → Approve → Deploy → Monitor → Update → Retire

Maker Permissions

To build agents, a maker typically needs:

• Appropriate Copilot Studio license/access
• Author permissions in the target Power Platform environment
• Membership in an Entra security group configured by administrators in the Power Platform admin center

Bring Your Own Agent (BYOA)

BYOA (Bring Your Own Agent) lets organizations integrate externally built agents — including Copilot Studio agents or custom platforms — into the Microsoft 365 Copilot experience or Teams via agent manifest registration. This is an integration pattern, not importing personal ChatGPT conversations or running models locally.

Worked Scenario: HR Policy Agent Rollout

An HR team builds a Copilot Studio agent grounded on SharePoint policy documents with an action to create a ServiceNow ticket. Before launch:

1. Security reviews knowledge source sensitivity labels
2. Admin submits agent for approval in M365 admin center
3. Teams app permission policy allows the agent for HR security group only
4. Power Platform admin verifies connector DLP permits ServiceNow
5. Post-launch, admins monitor usage in both admin centers

Common Exam Traps