Building Production-Ready Agents Microsoft Agent Framework

Devnot - Dotnet Konferansı 2026
07.05.2026
Sheraton Grand Ataşehir, İstanbul

Transcript

1. Building Production-Ready Agents with Microsoft Agent Framework Mert Metin #dotnetkonf26
2. MERT METİN Senior Software Engineer Blogcu Mühendis mertmtn.blogspot.com in/mrtmtn
3. Agenda Concepts Agents, Agent Orchestrations, Workflows Microsoft Agent Framework
4. Model call Single-agent with tools Multi-agent orchestration Complexity increases left to right Evaluate your scenario and choice right strategy
5. Agentic Loop Decide to use tools and plan the strategy. Think from prompt, instructions. Acting to call API, functions, RAG, other agents. Observe and evaluate the output of the action and make decision for restarting the loop.
7. AI Orchestration, coordinates multiple specialized agents within a unified system to achieve goals. Advantages: specialization, scalability, modularity, optimization, flexibility
8. Multi Agent Orchestration Patterns Sequential Concurrent
9. Group Chat Handoff Multi Agent Orchestration Patterns
10. Multi Agent Orchestration Patterns Magentic
11. Workflows Agents connect and work togetger to achieve a goal. Deterministic and predefined sequence of operations. Components are graph based.
13. Semantic Kernel LLM integration oriented framework. Use Cases: Content generation, chatting AutoGen Multi-agent LLM systems, provides orchestration mechanisims. Use Cases: Code Generation, Data Analysis + Unified Framework Merging
14. Why MAF? Open Source & Community Driven Interoperability: Supports A2A (Agent-to-Agent), MCP standards. LLM provider flexible - Without changing the code, swappable providers. Production-oriented: Middleware, caching, logging, dependency injection, telemetry Built-in state & session management Workflow orchestration (graph-based) Supporting Human-in-the-loop, checkpointing with in the workflows.
15. Agents in MAF Agent wrapped in AIAgent type which has persona and prompts. Simple agent creation: working agent in just a few lines of code. dotnet add package Microsoft.Agents.AI--version 1.3.0
16. ChatClientAgent : AIAgent Provides an "AIAgent" that connects directly to chat-completion services via IChatClient. name: Name of the AI Agent, instructions: Agent’s description with the ChatRole.System role
17. RunAsync Runs the agent with a user prompt. Returns response with AgentReponse type.
18. Structured AI Using generic type of AgentResponse<T>
19. RunAsync RunAsync<T>
20. Streaming Reponse Typing anmation effect with RunStreamingAsync Return type is IAsyncEnumerable<AgentResponseUpdate> which provides a stream of updates generated by the agent.
21. AgentSession AgentSession ensures agents have relevant context for long running dialogues. CreateSessionAsync creates a session object to preserve conversational context. Last two questions are testing turns to session object. short-term memory: Immediate context. Multi turn dialogue within a single execution. long-term memory: Persistent data for user preferences, past facts.
22. Middleware Build middleware method with Use()
23. Workflows in MAF dotnet add package Microsoft.Agents.AI.Workflows --version 1.3.0 A workflow connects edges and executors together into directed graphs and manages execution, routing messages.
24. AgentWorkflowBuilder Provides utility methods for constructing workflows composed of agents. Magentic Orchestration is not yet supported in C#.
25. Sample: ETL Workflow Extract, Transform, Load data with BuildSequential()
26. Sample: Legal Assistant BuildConcurrent() with two agents.
27. Executors: A work unit within the workflow AI Agents Business Logic
28. Edges : Data flow path between executors Define data flow path between executors in a workflow graph. It can include conditions to make routing decisions. Direct, Conditional, Fan-in, switch-case are supported with AddEdge()
29. Human-in-the-loop (HITL)
30. Demo: Spam Detection - Mail Processing Four Executors - Three Edges - Two Agents
31. Executors and conditional edges are builded with WorkflowBuilder.
34. WatchStreamAsync must begin iterating before TrySendMessageAsync is called, otherwise events emitted synchronously during processing will be missed.
35. Observability Integrating OpenTelemetry tracing into the agent - UseOpenTelemetry(source) dotnet add package OpenTelemetry --version 1.15.3 dotnet add package OpenTelemetry.Exporter.Console --version 1.15.3
36. Agent telemetry data
37. If you are still actively developing a project using SK or AutoGen You're reaching the complexity level of SK. So, you might consider migrating to MAF. If your project is stable or maintenance mode You don't have to switch to Microsoft Agent Framework. If you are starting a new project You need multi-agent orchestration, conversation history. So, you should start Microsoft Agent Framework. Migration Decisions
38. The takeaway If you're starting a new project, start with Agent Framework. LLMs are not enough. Agent based solutions are evolving. Agents bring intelligence, workflows brings structure. They form a coordinated AI system together.