Advanced Copilot Studio – two agent environments, multi-agent systems and operations (AICS2) (AICS2)

Course content:

Hide details

• Two generations of agents and why it matters
1. Standard (classic) and new agent environments side by side: why Microsoft rebuilt the environments in 2026 and what this changes for agent builders
2. The shift from conversation to agent: from explicit topics and branching to natural-language agent descriptions and improved orchestration
3. What “production-ready preview” means, coexistence of both environments and the fact that agents cannot be converted between them
4. Unified new-environment workspace: Build, Preview, Evaluate and Monitor, with credit billing already applying during building and testing

• The new environment in depth
1. Agent building blocks: instructions as a contract, knowledge, tools, model and connected agents
2. Skills: modular reusable instruction blocks, Markdown export and sharing between agents; replacing the standalone prompt node
3. Hooks: automatic workflow execution at points in the agent lifecycle; replacing topics and triggers
4. Memory and native code interpreter: when to enable persistent per-user context and what the computational tool is for
5. Improved orchestration runtime: works across all agents, requires no manual configuration and provides deeper access to Microsoft 365 data

• The classic environment when the new one is not enough
1. Topics, triggers and deterministic branching; condition groups for clearer logic
2. Generative orchestration in the classic environment and when to enforce controlled topic flow
3. What only the classic environment currently supports: custom entities, manual and custom authentication, SharePoint, WhatsApp and Dynamics 365 Contact Center channels, and Azure AI Search knowledge bases
4. Diagnosing incorrect tool selection and troubleshooting responses

• Difference map and environment selection
1. Capability comparison by category: tools, knowledge, channels, authentication, deployment and computer use (available in the classic environment and currently planned for the new one)
2. Decision criteria based on the use case, required channels, security and costs
3. Gaps in the new environment and how to work around them or temporarily replace them with a classic agent
4. Model selection: GPT-5.5 Chat, Claude Sonnet 5, reasoning models and experimental models; quality, latency, price and credit consumption

• Knowledge over company data
1. SharePoint within the tenant: grounding through the tenant graph, semantic search and metadata-based filtering
2. Dataverse tables, including unstructured columns, Azure SQL and Microsoft Graph connectors
3. External and analytical knowledge: Foundry IQ, Fabric IQ and Work IQ, and how they differ from Azure AI Search, which is not supported by the new environment
4. Why RAG returns the wrong passages: document structure and annotation, chunking, permission enforcement and retrieval quality measurement

• Workshop
1. Build the same agent in both environments over one knowledge base
2. Compare orchestration, response quality and development effort
3. Decide which environment is more suitable for the given use case and justify the decision

• Extending agents with custom systems
1. Connectors and tool groups; create a custom connector from an OpenAPI description, authentication and permission inheritance
2. Model Context Protocol: connect an MCP server as an agent tool, when to choose MCP over a classic connector, server certification and security implications
3. Workflow as an agent tool: multi-step processes, prompt node, Microsoft 365 Copilot node, and extracting values and tables from documents
4. Computer use: automating applications without APIs, cloud PCs through Windows 365, credential storage and audit logging; availability depends on the selected environment

• Multi-agent systems
1. Connected agents, nested agents and connections to data agents (Foundry, Microsoft Fabric)
2. Collaboration protocols: agent-to-agent and agents built on the Microsoft 365 Agents SDK
3. Design patterns: entry agent as a router, domain-specific specialists and shared tools
4. When a multi-agent solution is counterproductive and when a single agent is enough

• Autonomous agents and human-in-the-loop
1. Event- and schedule-based execution: triggers, hooks and agent flows
2. Long-running and asynchronous processes that exceed the two-minute limit
3. Requesting information or approval during execution and designing safe points for human intervention

• Deployment, versioning and evaluation
1. Solutions, environment variables and connection references; managed solutions and pipelines between Dev, Test and Prod
2. Developer workflow: Visual Studio Code extension, component collections and agent versioning in Git
3. Test sets and evaluation methods, multi-turn tests, comparing two agent versions and integrating evaluations into CI/CD
4. Content safety testing and early detection of regressions between versions

• Operations: monitoring, security and costs
1. Analytics and monitoring: user query topics, user satisfaction, custom metrics and telemetry export to Application Insights
2. Identity and governance: automatically created Microsoft Entra Agent IDs, Microsoft Agent 365, agent inventory and DLP policies for individual agents
3. Prompt injection and data leakage in production: defense-in-depth layers, handling sensitive data and audit trails
4. Costs and capacity: credit billing during both building and operation, cost estimation before deployment and optimization of expensive steps

• Workshop
1. Multi-agent solution with an autonomous trigger and human-in-the-loop
2. Custom test set and deployment to another environment through a pipeline
3. Choose the appropriate environment for each individual agent and justify the decision