Copilot Studio

Projects such as OpenAI's 'Operator', Anthropic's 'Computer Use', and Google's 'Project Mariner' are all exploring the application forms of AI agents in different scenarios.

OpenAI的Function Calling机制、Anthropic的Tool Use API以及开源框架LangChain、AutoGen均属于工具调用型智能体范畴

微软Copilot Studio将Agent构建能力直接开放给企业用户

Microsoft's Azure AI Agent Service and Amazon Bedrock Agents are also important players in the managed Agent space.

Copilot Studio has a limit on the number of tools available to an agent, approximately 70, so Business Central provides Dynamic Tool Mode.

Agent building in Copilot Studio revolves around three core components: the orchestrator, triggers, and tools.

Copilot Studio supports more than 1,500 Power Platform connectors, meaning trigger signals can be captured from almost any enterprise system.

Copilot Studio provides a Request for Information action to pause the Flow and request human review, as well as approval scenarios (including multi-level approval and AI approval in preview) covering approve/reject decisions.

Flow Checker lists all errors, and a Flow with errors cannot be published; after publishing, the Activity tab shows each run and its failures, while the Analytics tab shows run counts, failure rates, and trends.

Variables in Copilot Studio have four scopes: Topic variables (existing only within the Topic they belong to, the default), global variables (shared by all Topics), system variables (created by Copilot Studio, containing user and conversation context), and environment variables (coming from Power Platform, read-only in Copilot Studio).

Copilot Studio's built-in knowledge sources include public websites, SharePoint, Dataverse tables, and uploaded files, and it also supports connectors and Azure AI Search as knowledge sources.

Power Platform Connectors are real-time API bridges to thousands of systems; as tools they read/write data or trigger actions at runtime, and as knowledge sources (a preview feature) they can answer based on real-time data without copying it.

Microsoft has launched the AB-620 certification exam in the Copilot Studio ecosystem, aimed at professional developers and advanced builders who build enterprise-grade AI Agents.

Copilot Studio offers three authentication options: no authentication, Microsoft authentication (connecting to channels such as Teams, Power Apps, and Microsoft 365 Copilot via Entra ID), and manual authentication (supporting Entra ID or any OAuth2 identity provider).

Component Collections is a mechanism in Copilot Studio for packaging related topics, tools, knowledge sources, and even sub-Agents for reuse, supporting permission control and able to be restricted to use by a single main Agent only.

Agent Flow is deterministic automation in Copilot Studio, composed of a trigger and at least one action, with each action consuming Copilot Studio capacity.

Agents in Copilot Studio automatically determine at runtime whether a particular MCP tool needs to be invoked based on user intent, and the process is transparent to end users.

Microsoft's agent framework, Copilot Studio, allows enterprises to customize the tool-calling boundaries and trigger conditions of agents.

Memory capability can be enabled in Copilot Studio, allowing the agent to learn how a user operates.

Maker-provided credentials is a connector authentication mode in Copilot Studio that, once enabled, causes all anonymous users' actions to be executed under the identity of the maker.