Building and Deploying AI Agents in Microsoft Copilot Studio: Copilot Chat, Standard, and GitHub Copilot Harnesses

This course is intended for:

• AI solution architects and consultants, who design Copilot Studio implementations and must deliberately choose the harness, orchestration pattern, and cost model for a specific process.
• Experienced Power Platform developers and low-code makers, who have already built an agent using one approach and want to learn the others and how to combine them.
• IT professionals and Microsoft 365 and Power Platform administrators, responsible for environments, permissions, data loss prevention (DLP) policies, publishing, solution lifecycle, and capacity management.
• Technical leads and AI product owners, who are responsible for moving pilots into production – including governance, quality evaluation, and budget.
• Trainers and Center of Excellence (CoE) teams, who build internal agent development standards for their organization.

After completing this course, participants will be able to:

• Choose the right harness – decide between the Copilot Chat, standard, and GitHub Copilot harnesses based on requirements rather than habit, and justify that decision to the business.
• Work across the entire platform – build agents in Agent Builder, in the agentic reasoning loop, and in the classic topic-based model, learning the strengths and weaknesses of each approach.
• Master Instructions and Skills – apply the “always vs. sometimes” principle and progressive disclosure, and write descriptions that actually invoke the right component.
• Connect Knowledge, Tools, and Workflows – connect the agent to enterprise data (SharePoint, Dataverse, connectors, MCP) and automate event-driven processes with the new Workflows experience.
• Build multi-agent systems – design a team of specialists (child agents and connected agents) using the hub-and-spoke pattern, with deliberate context passing.
• Build autonomous agents – run an agent from an event trigger and hand the result over to a human as an interactive Adaptive Card in Microsoft Teams.
• Apply responsible AI and governance – layer three content-blocking mechanisms, write a complete AI disclosure, and place the agent in the appropriate risk zone.
• Implement ALM and measurable quality – move a solution between environments using solutions, environment variables, and connection references, and prove improvements with evaluations rather than gut feeling.
• Control costs (AI FinOps) – understand Copilot Credits, price a conversation, and learn why fewer turns beat a cheaper model.

In practice, you will leave the course able to design, build, secure, deploy, and cost an agent in any of the three approaches – rather than knowing only one of them

Module 1 – Microsoft 365 Copilot and the Copilot Chat Harness

• Why AI deployments stall; the operating model and the Frontier Firm; Microsoft 365 E7 licensing
• What an agent is: orchestrator, model, memory, knowledge, tools, skills, and governance; the agent spectrum and five types of agents
• The Microsoft agent ecosystem; choosing the harness before you build; orchestration patterns and structured reasoning
• Microsoft 365 Copilot: Work IQ, memory, model selection, Agent Store, and the Researcher and Analyst agents
• Copilot Chat harness: orchestration loop, component model, first-party capabilities
• Four ways to build: Agent Builder, Copilot Studio, Microsoft 365 Agents Toolkit, Work IQ Dev Tools; MCP servers and MCP Apps
• Labs: M365 Copilot and Frontier Agents and Build AI Assistants with Agent Builder

Module 2 – GitHub Copilot Harness: Instructions and Skills

• Why the agentic reasoning loop (thought – action – observe – decide) closes the gaps of classic orchestration
• Component model: instructions, knowledge, tools, skills, memory, agent sandbox, connected agents
• Knowledge retrieval: search, ranking, and downloading full files into the sandbox; per-user and per-agent memory
• Migrating from the standard harness: component mapping, rebuilding instead of one-to-one porting
• Instructions: what always applies vs. what descriptions decide; names, inputs, outputs, and dynamic chaining
• Skills: anatomy of a SKILL.md package, progressive disclosure, the determinism dial from goal to Python script
• Labs: The Agentic Reasoning Loop and Deep Dive: Skills

Module 3 – GitHub Copilot Harness: Knowledge, Tools, and Workflows

• Knowledge sources: SharePoint and OneDrive, files, Dataverse, public websites, Microsoft 365 Copilot connectors, Azure AI Search
• Retrieval pipeline: query rewriting, retrieval, summarization – with validation at every step; security trimming
• Seven types of tools: prebuilt and custom connectors, REST APIs, MCP servers, agent flows, prompts, computer use
• Connector vs. REST API vs. MCP decision; RPA vs. computer use; combining multiple tools in a single agent
• Workflows: event-driven automation, trigger and payload, building blocks, determinism vs. AI on a single canvas
• Agent node: inline vs. referenced; human in the loop; testing nodes, Activity and Monitor
• Labs: Copilot Studio Tools and Workflows

Module 4 – Standard Harness: Orchestration and Multi-Agent Systems

• When the standard harness is the right choice: mandated predictability, channel constraints, required message wording
• Component model: topics, knowledge, tools, variables, agents, channels, analytics
• Topics: triggers and nodes; variables and their scope (topic, global, cross-session)
• Multi-agent systems: child agents vs. connected agents; hub-and-spoke, pipeline, and collaborative patterns
• Passing context between agents and instruction best practices for an agent team
• Classic vs. generative orchestration; inputs, outputs, and dynamic chaining; debugging incorrect routing
• Labs: Create the Agent, Adding the Child Agent, Adding the Connected Agent

Module 5 – Standard Harness: Event Triggers and Adaptive Cards

• Topic trigger vs. event trigger; event sources in Microsoft 365, Dataverse, and on a schedule (recurrence)
• Payload: everything the agent knows at the start; filtering on the trigger instead of in the flow
• Where autonomy pays off and where it is just noise; identity, connection references, DLP, and publishing
• Adaptive Cards as a data structure: body, elements, actions, and schema version
• Live data binding – one card definition for every record; designing and debugging cards
• Labs: Automating Event Triggers and Notify a Teams Channel with an Adaptive Card

Module 6 – Operations: Responsible AI, ALM, and Cost Management

• The six Microsoft Responsible AI principles; distinguishing AI safety, security, and governance
• Three independent content-blocking mechanisms; complete AI disclosure; prompt injection and Agent Runtime Protection
• The zone model (green, yellow, red): capabilities vs. potential blast radius
• Why ALM for AI is a loop, not a line; solutions, environment variables, and connection references
• Pipelines in Power Platform, source control, and component collections; analytics vs. evaluations; test coverage zones
• AI FinOps: Copilot Credits, conversation pricing, purchasing models, three layers of control, and the “rediscovery tax”
• Labs: Responsible AI and content moderation (classic or new experience version to choose from) and Monitor Performance and Evaluate Agent Quality

• Hands-on experience with Microsoft 365 services (Teams, SharePoint, Outlook) at an advanced user level.

• Basic knowledge of Microsoft Power Platform: environments, solutions, and connectors.

• General understanding of artificial intelligence and large language model (LLM) concepts – tokens, context window, retrieval-augmented generation (RAG).

• Prior experience building at least one agent or flow (Power Automate, Copilot Studio, Agent Builder) is an advantage.

• No programming experience is required – the course follows a low-code / no-code approach.

• Completion of AI-901 or PL-900, or equivalent knowledge, is an advantage.

•  manual in electronic form available on the platform: https://www.altkomakademia.pl/
• access to Altkom Akademia's student portal

Training method:

• Lecture: 30%

• Workshops: 25%

• Labs:45%

• Training: English

• Materials: English