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.

For scenarios with relatively fixed inputs and outputs such as "extracting insights from emails," building a workflow with branching logic in Copilot Studio is more reliable than letting the Agent act freely.

Using the Perform an unbound action step in a Copilot Studio workflow can call a custom API whose binding type is global; choosing the wrong binding type will cause the action not to be found in the workflow.

The Copilot Studio workflow can be used to implement an unattended, automatically confirming closed loop for Dynamics 365 purchase orders: it ends silently when all orders succeed, and on failure it wakes the agent to investigate and pushes a notification to Teams.

The custom API in this demo requires manually entering 3 request parameters (purchase_order_numbers, stop_on_first_failure, company_id) and 4 response properties (confirm_count, failed_count, results_json, company), and there is currently no capability to generate them automatically from the class.

Placeholder content in the prompt goes directly into the agent's context window; if results_json is large, it may crowd out reasoning space or increase token consumption, so for batch failure scenarios it is recommended to truncate or summarize first.

Downgrading the agent from the main flow to an exception handler, i.e., the workflow handles routine actions and the agent is woken only on failure branches, can reduce token consumption and response latency.

Generative orchestration in Copilot Studio relies on the reasoning capabilities of large language models to dynamically decide which plugin to call and how to respond based on system instructions, knowledge sources and available tools, making it suitable for open-ended questions but with lower behavioral predictability.

Zakaria suggests splitting instructions into five blocks: who the agent is, what it does, what it refuses, how it answers, and what to do when it doesn't know.

When fine-grained model selection or large-scale cost control is needed, Azure AI Foundry is often more suitable than Copilot Studio, and in large groups the two often coexist.

The six-step method will not make the agent incapable of making mistakes; its value lies in making failures visible and fixable.

Zakaria Khchiche is a freelance Data & AI Tech Lead in Paris, has worked on 20 generative AI use cases at SUEZ, and has delivered more than 200 AI agents.

In Copilot Studio practice, a common hybrid strategy is to hand consultation, retrieval and explanation needs to generative orchestration, while hardening high-risk steps such as identity verification, amount confirmation and compliance notices into hand-crafted topics or controlled flows.

Hand-crafted topics are like flowcharts: developers preset trigger phrases, branching conditions and reply templates, and the agent must follow the established path, falling back to a default answer when out of scope.

Power Platform is Microsoft's set of low-code/automation business platforms, whose core components include Power Apps, Power Automate, Power BI and Copilot Studio.

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.