microsoft-agent-framework

name

microsoft-agent-framework

description

Master skill for Microsoft Foundry + Microsoft Agent Framework (Python) development — agent patterns, deployment path selection (prompt agent via azure-ai-projects SDK, hosted agent via azd/container/ZIP), the Foundry Toolkit for VS Code, publishing to M365/Teams via the Agent 365 CLI, evaluation, tools, and .NET↔Python API bridging. Delegates to task-oriented leaf skills. Use when working on any Foundry or Agent Framework project in this repo. Triggers on: "agent framework", "maf", "foundry", "magentic", "workflow builder", "hosted agent", "prompt agent", "foundry agent", "agent deployment", "deploy to foundry", "file search", "azd ai agent", "foundry toolkit", "agent 365", "publish to teams", "agent eval", "work iq", "workiq", "m365 data", "work iq setup", "work iq error", "workiq troubleshooting", "workiq 403", "wam error", "workiq ebusy", "run demo error", "unicodeencodeerror", "charmap codec", "cp1252", "emoji crash", "servers won't shut down", "ctrl_break", "process group", "health check timeout", "address already in use", "function tool not supported", "a2a run error", "declarative workflow", "csdl", "power fx", "powerfx", "conditiongroup", "foreach", "set variable", "if else workflow", "workflow empty canvas", "workflow routing not running", "invokeazureagent", "sendactivity", "workflowagentdefinition", "portal workflow", "visual designer", "agent_reference".

microsoft-agent-framework — Agent Framework Router

Routes Agent Framework requests to the appropriate leaf skill. Do not implement directly — delegate immediately.

Delegate Map

Request type	Leaf skill to invoke
Add/edit agents, change workflow topology, orchestration (code-first WorkflowBuilder SDK)	agent-framework-workflow
Build a Magentic team against Foundry; "No tool call found for function call output"; max-reset loops; Chat-Completions vs Responses client choice	agent-framework-magentic-foundry
Build/publish a Foundry portal workflow (CSDL YAML); empty-canvas or routing-never-runs bugs; ConditionGroup / Foreach dialect	foundry-workflow-declarative
Write/debug Power Fx expressions in a portal workflow (conditions, values, .-on-Text errors, scopes)	foundry-workflow-powerfx
Design workflow architectures, choose patterns, reference topologies	agent-framework-orchestration
Coordinate .NET ↔ Python contracts, API endpoints, DTOs	agent-framework-api-bridge
Deploy a prompt agent (SDK) OR a hosted agent (azd / container / ZIP)	agent-framework-deploy
Build / debug / eval an agent from the VS Code Foundry Toolkit	foundry-toolkit
Publish an agent to Microsoft 365 / Teams; govern its Entra identity	agent-365-cli
Eval datasets, batch eval, prompt optimization	agent-framework-eval
Connect an agent to live M365 data with Work IQ; provision / consent / bill Work IQ	work-iq-setup
Diagnose a failing Work IQ install / auth / consent / billing / query	work-iq-troubleshooting
Diagnose a demo that fails when run locally (encoding crash, servers won't start / shut down, port in use, model tool support)	agent-framework-run-troubleshooting

Work IQ vs Agent 365 CLI: Work IQ reads live M365 data into an agent (mail, calendar, Teams); the Agent 365 CLI publishes an agent out to Teams/M365 as a registered identity. They are complementary, not alternatives.

Deployment path by agent type (do not default to raw az rest):

• Prompt agent (instructions + model + tools) → azure-ai-projects 2.x SDK: project.agents.create_version(PromptAgentDefinition(...)). See agent-framework-deploy.md.
• Hosted agent (your own code) → azd ai agent (azure.ai.agents extension: azd ext install azure.ai.agents) or the SDK ZIP path. See agent-framework-deploy.md.
• Publish to Teams/M365 → Agent 365 CLI (a365). See agent-365-cli.md.

Shared References

• Stack Versions — Current package versions (core 1.9.0 stack), API signatures, model client + orchestration decision tables
• Code Patterns — FoundryChatClient, MagenticBuilder, WorkflowBuilder, non-Foundry models
• Magentic on Foundry — Magentic teams MUST use a Chat Completions client (not Responses); no Claude in the loop; one client per agent; single task message