Microsoft

Tools are what an agent invokes beyond answering from knowledge. They call APIs, run automated flows, reach external services and fetch real time data, and the orchestrator picks one from the conversation context and the tool's description. Connections and connectors carry the credentials. An agent can be extended with Model Context Protocol so every tool and resource a connected MCP server publishes becomes available automatically, with changes on that server reflected without rework.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/tools-overviewread 2026-09-17

Generative orchestration lets a custom agent answer with the relevant topics, tools, other agents and knowledge sources, fill inputs from earlier conversation history, chain several capabilities together in one response, and act on event triggers, with a triggered agent choosing which tool, topic or agent to invoke. Deterministic work sits beside it as agent flows, and other agents are attached as connected agents and managed from the agent's own Agents page.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/faqs-generative-orchestrationread 2026-09-17

Knowledge sources are content the maker attaches at design time so every user gets the same grounding. The Add knowledge dialog offers featured sources (public websites, SharePoint, OneDrive for Business, Salesforce, ServiceNow, Azure SQL), advanced sources (Azure DevOps wiki and work items, custom connector, enterprise websites) and direct upload of PDF, Word, Excel and PowerPoint files.

Dataverse tables, Azure AI Search, Bing Custom Search, real time connectors and file groups can also be used, and sources can be managed and edited. New knowledge goes into an index without retraining the model, and a file an end user uploads is handled as an attachment to the conversation rather than as a knowledge source.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/knowledge-sources-overviewread 2026-09-17

Human supervision is a configurable setting. When the agent needs confirmation or more information mid execution, it escalates to the named reviewer, sending a review request by Outlook email or showing an inline review card in the activity panel, and a published table shows which underlying models support the feature. Microsoft also covers administration of computer use and its troubleshooting and limitations, and warns explicitly that these agents can meet prompt injection attempts and should run in trusted isolated environments with validation before instructions are executed.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/human-supervision-computer-useread 2026-09-17

Copilot Studio is covered by Microsoft's standards certifications and regulatory compliance programs, and its customer facing controls include geographic data residency, data loss prevention, environment routing, regional customization, an agent runtime protection status view for makers and an automatic security scan. Where an organization onboards Microsoft Agent 365, Copilot Studio agents are represented as identities in Microsoft Entra and governed with Conditional Access, role based and attribute based access control and access governance workflows, in addition to the existing Power Platform and Microsoft 365 mechanisms.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/security-and-governanceread 2026-09-17

After publishing, the Monitor tab reports conversation sessions, user engagement with message counts and session duration, credit consumption over a chosen period, thumbs up and thumbs down reactions with comments on specific responses, and response quality indicators over time, filterable by time range. Errors can be reviewed in their own view. Computer use runs have their own monitoring, and where Agent 365 is onboarded it carries centralized observability and policy enforcement across agents rather than only where each was built.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/analytics-overviewread 2026-09-17

Memory gives an agent a separate store per user, held in a dedicated folder in Microsoft managed storage, with a lifecycle of capture, store and apply. Preferences and context from a conversation are written as files and read back on later interactions. The maker turns it on or off per agent, and a user's memories are private to that user and invisible to the maker. Each user can review and remove what the agent remembers, memories are deleted automatically after 28 days without interaction, and turning the feature off stops use without deleting.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/memory-overviewread 2026-09-17

In Copilot Studio, geographic data residency covers where data is stored, processed and managed, tied to the environment's geography. Microsoft also covers data locations, the security side of residency, and deliberately moving data across regions when generative features are needed outside the geography or capacity is constrained. Model availability is published per region in tables on the model selection page, with a warning that data in an experimental model may be processed outside the organization's geographical boundary. Virtual network support is also offered.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/geo-data-residencyread 2026-09-17

The Agents page in Copilot Studio offers a set of agent templates, each a complete agent a maker creates from and then customizes. IT Helpdesk, Safe Travels, Weather, Website Q&A and Sustainability Insights are custom agents, and Benefits is a Microsoft 365 Copilot agent, each with its own documentation page. The catalog sits inside the product and can be browsed, each template works as soon as it is chosen, and the set is small.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/template-fundamentalsread 2026-09-17

Event triggers let an agent act on its own rather than waiting for a person. A chosen event generates a payload, a connector delivers it, and the agent executes against its instructions and the payload's variables. Topic triggers, by contrast, need a user's phrase. In Microsoft's example an agent starts when a new entry lands in a Dataverse table, and agents also reach users through conversational channels and the Microsoft 365 apps.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/authoring-triggers-aboutread 2026-09-17

The maker chooses the primary model for an agent's generative orchestration from a dropdown, with separate settings for deep reasoning, for generative responses and for the prompt builder. An external model can be chosen as the primary model, and Microsoft publishes tables of model availability by region and by scope. Models carry tags for their status, experimental ones are marked as not recommended for production with a limitations page, and makers can continue on a model after it is retired.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/authoring-select-agent-modelread 2026-09-17

The Microsoft 365 Agents SDK builds agents in code in C# and .NET, JavaScript on Node.js, and Python, and such an agent is attached to a Copilot Studio agent by its messaging endpoint so the platform can call it. The platform is itself callable. The Power Platform REST API triggers agent evaluations programmatically for release validation, regression testing and CI/CD pipelines, and separate admin APIs delete, quarantine and reassign ownership of agents. Model Context Protocol brings external tools and resources in.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/add-agent-microsoft-365-agents-sdk-agentread 2026-09-17

Evaluations are named test sets. A conversation is a test case with user messages and optionally expected responses, created by hand, generated with AI, or uploaded from CSV, grouped into an evaluation with a test method, and run to produce scored results that track quality across changes and validate behavior before publishing. Evaluations run under a chosen authenticated user profile so the agent's tools and connections are exercised. The Power Platform REST API runs them programmatically against predefined test sets for release validation, regression testing and CI/CD.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/agents-experience/analytics-agent-evaluation-introread 2026-09-17

Computer use is a tool an agent adds from its Tools page. It interacts with a Windows computer through a virtual mouse and keyboard, working across websites and desktop applications by selecting buttons, choosing menus and typing into fields, described in natural language rather than code, for work where no API exists.

It is powered by a computer using agent model combining vision with reasoning, which Microsoft says adapts when buttons or screens change rather than breaking. Microsoft also covers where runs execute, human supervision, monitoring, administration and limitations for computer use.

Sourcelearn.microsoft.com/en-us/microsoft-copilot-studio/computer-useread 2026-09-17