Much like a high-performing corporation relies on specialized departments, we must transition from a single-executor model to a\u00a0Collaborative Multi-Agent Network.<\/strong><\/p>\n The\u00a0Microsoft Agent Framework<\/strong>\u00a0is designed to address this paradigm shift, offering a unified, observable platform that empowers developers to achieve two core value propositions:<\/p>\n Each Agent serves as a specialized, pluggable, and independently operating execution unit, underpinned by three critical pillars of intelligence:<\/p>\n Developers retain the flexibility to utilize leading model providers, including Azure OpenAI, OpenAI, Azure AI Foundry or local models, to customize and build these high-performance Agent primitives.<\/p>\n The Workflow feature is the flagship capability of the Microsoft Agent Framework, elevating orchestration from simple linear flow to a dynamic collaboration graph. It grants the system advanced architectural abilities:<\/p>\n Crucially, Workflow definitions are not limited to Agent connections; they can integrate seamlessly with existing business logic and method executors, providing maximum flexibility for complex process integration.<\/p>\n Drawing on the GitHub Models examples, we demonstrate how to leverage the Workflow component to enforce structure, parallelism, and dynamic decision-making in enterprise applications.<\/p>\n In sophisticated production scenarios, these patterns are frequently layered: for instance, a Concurrent search and summarization phase followed by a Conditional branch that routes the result to either automatic publishing or a Sequential Human-in-the-Loop review process.<\/p>\n For complex multi-agent systems,\u00a0Observability<\/strong>\u00a0is non-negotiable. The Microsoft Agent Framework offers an exceptional developer experience through the built-in\u00a0DevUI<\/strong>, providing real-time visualization, interaction tracking, and performance monitoring for your orchestration layer.<\/p>\n The following simplified code demonstrates the key steps to enable this capability in your project (see project\u00a0 By pairing the DevUI’s visual execution path with APM trace data, you gain the ability to rapidly diagnose latency bottlenecks, pinpoint failures, and ensure full control over your complex AI system.<\/p>\n Multi-Agent Orchestration represents the future of complex AI architecture. We encourage you to delve deeper into the Microsoft Agent Framework to master these powerful capabilities.<\/p>\n Here is a curated list of resources to accelerate your journey to becoming an Agent Architect:<\/p>\n Microsoft Agent Framework GitHub Repo:<\/strong>\u00a0https:\/\/github.com\/microsoft\/agent-framework<\/a><\/p>\n<\/li>\n Microsoft Agent Framework Workflow official sample:<\/strong>\u00a0https:\/\/github.com\/microsoft\/agent-framework\/tree\/main\/python\/samples\/getting_started\/workflows<\/a><\/p>\n<\/li>\nScenario 1: Architecting Professionalized AI Agent Units<\/h3>\n
\n
Scenario 2: Dynamic Coordination via Workflow Orchestration<\/h3>\n
\n
Deeper Dive: Workflow Patterns<\/h2>\n
1. Sequential: Enforcing Structured Data Flow<\/h3>\n
![\"wf01](\"https:\/\/devblogs.microsoft.com\/semantic-kernel\/wp-content\/uploads\/sites\/78\/2025\/10\/wf01.png\")
<\/a><\/p>\n\n
to_reviewer_result<\/code>) at intermediate nodes for data formatting, validation, or status logging, thereby establishing critical checkpoints.<\/li>\n<\/ul>\n# Linear flow: Agent1 -> Agent2 -> Agent3\r\n\r\nworkflow = (\r\n\tWorkflowBuilder()\r\n\t.set_start_executor(agent1)\r\n\t.add_edge(agent1, agent2)\r\n\t.add_edge(agent2, agent3)\r\n\t.build()\r\n)<\/code><\/pre>\n\n
2. Concurrent: Achieving High-Throughput Fan-out\/Fan-in<\/h3>\n
<\/a><\/p>\n\n
aggregate_results_function<\/code>), where custom logic must be implemented to reconcile multi-branch returns, often via voting mechanisms, weighted consolidation, or priority-based selection.<\/li>\n<\/ul>\nworkflow = (\r\n\tConcurrentBuilder()\r\n\t.participants([agentA, agentB, agentC])\r\n\t.build()\r\n)<\/code><\/pre>\n\n
3. Conditional: State-Based Dynamic Decisioning<\/h3>\n
<\/a><\/p>\n\n
selection_func<\/code>). It receives the parsed intermediate data (e.g.,\u00a0ReviewResult<\/code>) and returns a list of target executor IDs, enabling not just single-path routing but also complex logic where a single data item can branch into multiple parallel paths.<\/li>\n<\/ul>\ndef select_targets(review, targets):\r\n\thandle_id, save_id = targets\r\n\treturn [save_id] if review.review_result == \"Yes\" else [handle_id]\r\n\r\nworkflow = (\r\n\tWorkflowBuilder()\r\n\t.set_start_executor(evangelist_executor)\r\n\t.add_edge(evangelist_executor, reviewer_executor)\r\n\t.add_edge(reviewer_executor, to_reviewer_result)\r\n\t.add_multi_selection_edge_group(to_reviewer_result, [handle_review, save_draft], selection_func=select_targets)\r\n\t.build()\r\n)<\/code><\/pre>\n\n
Production-Grade Observability: Harnessing DevUI and Tracing<\/h2>\n
main.py<\/code>):<\/p>\n\n
# Transform and selection function example\r\n@executor(id=\"to_reviewer_result\")\r\nasync def to_reviewer_result(response, ctx):\r\n\tparsed = ReviewAgent.model_validate_json(response.agent_run_response.text)\r\n\tawait ctx.send_message(ReviewResult(parsed.review_result, parsed.reason, parsed.draft_content))\r\n\r\ndef select_targets(review: ReviewResult, target_ids: list[str]) -> list[str]:\r\n\thandle_id, save_id = target_ids\r\n\treturn [save_id] if review.review_result == \"Yes\" else [handle_id]\r\n\r\n# Build executors and connect them\r\nevangelist_executor = AgentExecutor(evangelist_agent, id=\"evangelist_agent\")\r\nreviewer_executor = AgentExecutor(reviewer_agent, id=\"reviewer_agent\")\r\npublisher_executor = AgentExecutor(publisher_agent, id=\"publisher_agent\")\r\n\r\nworkflow = (\r\n\tWorkflowBuilder()\r\n\t.set_start_executor(evangelist_executor)\r\n\t.add_edge(evangelist_executor, to_evangelist_content_result)\r\n\t.add_edge(to_evangelist_content_result, reviewer_executor)\r\n\t.add_edge(reviewer_executor, to_reviewer_result)\r\n\t.add_multi_selection_edge_group(to_reviewer_result, [handle_review, save_draft], selection_func=select_targets)\r\n\t.add_edge(save_draft, publisher_executor)\r\n\t.build()\r\n)<\/code><\/pre>\n\n
main.py<\/code>)<\/li>\n<\/ol>\nfrom agent_framework.devui import serve\r\n\r\ndef main():\r\n\tserve(entities=[workflow], port=8090, auto_open=True, tracing_enabled=True)\r\n\r\nif __name__ == \"__main__\":\r\n\tmain()<\/code><\/pre>\nImplementing End-to-End Tracing<\/h3>\n
<\/a>\nWhen deploying multi-agent workflows to production or CI environments, robust tracing and monitoring are essential. To ensure high observability, you must confirm the following:<\/p>\n\n
.env<\/code> prior to start up.<\/li>\ntracing_enabled<\/code>\u00a0to\u00a0True<\/code>\u00a0and configure an\u00a0OpenTelemetry Protocol (OTLP)<\/strong>\u00a0exporter. This enables the complete execution call chain (Trace) to be exported to an APM\/Trace platform (e.g., Azure Monitor, Jaeger).<\/li>\nNext Steps: Resources for the Agent Architect<\/h2>\n
\n